# External Memory at Work: What Empirical Research Actually Establishes ## Executive summary The uploaded brief supplies a concrete research question despite the surrounding request describing the topic as unspecified: **what empirical research establishes about the benefits and costs of external memory, personal knowledge management, and shared knowledge systems for remembering, retrieving, deciding, and acting in real work, as of October 3, 2026**. It explicitly asks that popular “Second Brain” frameworks be audited rather than assumed valid. [1](https://chatgpt.com/g/g-p-6ac129ef43c08191acf988f750b7a6da/c/6ac13845-ef28-83ed-a64f-2ab6d52a698f) Citation style was not specified, so this report uses inline source citations and DOI links in the study table. The central conclusion is narrower—and more useful—than the standard “your brain is for having ideas, not holding them” story: > **Externalization is reliably useful when it converts a fragile internal memory requirement into a dependable future cue or a readily retrievable artifact. It is not reliably useful merely because information has been captured.** The best-supported effects concern **prospective memory and immediate task performance**. External reminders improve remembering to perform future actions, including in laboratory tasks and a 48-hour email task; people use reminders more as memory demands increase. But offloading is not free: people can overuse reminders relative to their objective need, externalization can weaken learning of the offloaded intention, and performance can deteriorate when a previously available aid disappears. Metacognitive feedback can improve calibration about when to offload. [2](https://link.springer.com/article/10.3758/s13421-025-01708-x) For **retrospective memory and learning**, the result depends heavily on *what is being optimized*. Saving information can release interference and improve learning of subsequently encountered material when the external store is believed reliable, but activities that encourage people to treat the external record as a substitute for encoding can reduce what they internally remember. Photographing museum objects impaired some later memory in Henkel's experiments, while Storm and Stone found that reliably saving one file could improve memory for subsequently studied information. These are not contradictory once “memory of the offloaded material” is separated from “capacity available for the next task.” [3](https://journals.sagepub.com/doi/10.1177/0956797614559285) For **note-taking**, the evidence does not justify the slogan “handwriting is superior.” Mueller and Oppenheimer's original three studies found worse conceptual performance with laptop notes and more verbatim transcription, but a direct replication reproduced the process differences without reproducing the immediate performance advantage of longhand. Other experiments suggest that generative processing—such as typing notes yourself rather than copy-pasting—can matter more than the physical medium. Structured note scaffolds can help in educational tasks, but transfer from classrooms to meetings, research, product decisions, or management remains largely untested. [4](https://www.psychologicalscience.org/journals/psychological-science/0956797614524581/) For **long-term remembering**, active retrieval is much better supported than passive storage. Retrieval-practice experiments show that repeatedly recalling learned material produces better delayed retention than repeated study. But this is evidence about *strengthening human memory*, not evidence that searching a note archive, opening a backlink, or having an AI retrieve a document generates the same effect. Treating the educational “testing effect” as proof for PKM search or backlinking would be a category error. [5](https://doi.org/10.1126/science.1152408) For **personal information management**, “keeping” and “finding” are distinct problems. Naturalistic studies show that users can navigate self-created folder structures quickly when those structures remain relatively shallow and folders remain manageable. In studies where folders and tags coexisted, users strongly preferred folders and performed worse with tags in a controlled retrieval task. Historical studies also found navigation remained much more common than search even when search improved. Shared-file research adds that large collections, multiple versions, deeper folders, larger sharing groups, elapsed time, and workload can all make retrieval worse. [6](https://onlinelibrary.wiley.com/doi/abs/10.1002/asi.21415) There is **no comparable empirical basis for the most distinctive claims of Zettelkasten, PARA, progressive summarization, backlinks, atomic notes, or Building a Second Brain**. Targeted searches located practitioner definitions, educational hybrids, historical descriptions, and many testimonials, but not independent controlled or longitudinal workplace evaluations showing that these frameworks outperform plausible alternatives. Forte's own materials define Second Brain and its CODE/PARA methods as productivity methodologies; that establishes what the framework claims, not whether those claims are causally true. [7](https://fortelabs.com/) At the team level, the evidence becomes stronger again. **Transactive memory**—a team knowing who knows what and coordinating access to distributed expertise—has experimental and longitudinal support. Teams trained together can outperform newly assembled teams, and a longitudinal study of 64 MBA consulting teams found more developed transactive-memory systems associated with team performance and viability. This construct is not equivalent to a shared wiki: the critical resource is partly *meta-knowledge about expertise*, not merely documents. [8](https://journals.sagepub.com/doi/10.1177/0146167295214009) **Organizational memory** also has real-world support, but the useful retention mechanism can be routines, standardization, roles, people, or organizational structure rather than a note repository. A 2025 study covering IT projects in 94 US federal agencies found that centralization and task standardization buffered project delays associated with workforce turnover. Earlier service-organization research showed organizational knowledge both accumulating and depreciating over time. [9](https://academic.oup.com/jpart/advance-article-pdf/doi/10.1093/jopart/muaf019/63489968/muaf019.pdf) **Decision logs have a narrower empirical foundation than proponents imply.** Recording and later reviewing what one believed before an outcome can reduce reconstructive hindsight bias. Davies found that access to records of pre-outcome thinking reduced hindsight bias in experimental settings. But simply creating a record was not sufficient in every condition, and there is no strong longitudinal evidence that keeping a decision journal causes executives or founders to make fewer repeated mistakes over years. [10](https://doi.org/10.1016%2F0749-5978%2887%2990005-7) AI expands what can be searched, summarized, and generated, but does not remove the old trade-offs. Professional field experiments show that generative AI can substantially improve speed and output quality on tasks that fall within model capability, while other research shows human–AI feedback loops can amplify bias. A small 2025 preprint on AI-assisted note-taking found that the most automated condition was easiest for participants but produced the weakest post-test performance, whereas intermediate assistance performed better. Direct longitudinal evidence that an AI layer turns a large personal archive into better decisions or execution is still missing. [11](https://aiinstitute.hbs.edu/navigating-the-jagged-technological-frontier/) The evidence can therefore be summarized as follows:
Question Best current answer Confidence Professional transfer
Do external reminders prevent forgotten intentions? **Yes, reliably** when the reminder is available and appropriately cued. [2](https://link.springer.com/article/10.3758/s13421-025-01708-x) High Moderate–high
Does external storage “free cognitive capacity”? **Sometimes**: reliable saving can reduce interference and aid subsequent learning/task performance. [3](https://journals.sagepub.com/doi/10.1177/0956797614559285) Moderate–high Moderate
Does offloading weaken internal memory? **Sometimes**, especially for the offloaded item/intention or when encoding effort is displaced. [12](https://www.psychologicalscience.org/journals/psychological-science/0956797613504438/) Moderate Moderate
Are people optimally calibrated about offloading? **No.** Decisions are strongly influenced by confidence; systematic over-offloading occurs in some paradigms. Feedback can improve calibration. [13](https://link.springer.com/article/10.1186/s41235-019-0195-y) High for lab paradigms Moderate
Is handwriting inherently better than typing? **No robust general advantage established.** Processing strategy matters more plausibly than medium alone. [4](https://www.psychologicalscience.org/journals/psychological-science/0956797614524581/) Moderate Low–moderate
Does active retrieval strengthen long-term memory? **Yes.** But file retrieval is not automatically retrieval practice. [5](https://doi.org/10.1126/science.1152408) High Moderate
Are tags/backlinks superior to folders? **No.** Tags have not consistently beaten folders; direct backlink evidence is notably absent. [14](https://onlinelibrary.wiley.com/doi/10.1002/asi.22906) Moderate for folders/tags; low for backlinks Moderate
Does storing more information create more value? **No evidence of monotonic value.** Larger collections and buried items can worsen refinding. [15](https://www.sciencedirect.com/org/science/article/abs/pii/S2050380620000137) Moderate–high High
Are Zettelkasten/PARA/Second Brain independently validated? **No independent framework-level validation located.** [7](https://fortelabs.com/) High confidence in evidence gap High
Do shared knowledge systems help teams? **Yes under some mechanisms**, particularly expertise location, coordination, standardized process, and continuity—not merely document accumulation. [16](https://pubsonline.informs.org/doi/abs/10.1287/mnsc.1040.0257) Moderate–high High
Do decision logs improve decisions? They can **preserve foresight and reduce hindsight reconstruction**; repeated-decision improvement is unproven. [10](https://doi.org/10.1016%2F0749-5978%2887%2990005-7) Moderate Moderate
Does AI make a huge archive operationally valuable? **Plausible but inadequately tested.** [17](https://arxiv.org/abs/2509.03392) Low Unknown
The practical implication is not “build a bigger Second Brain.” It is: **design external memory around future use**. A useful system needs reliable cues for commitments, low-cost refinding, selective capture, provenance for decisions, clear ownership in teams, deliberate internal learning when unaided knowledge matters, and deletion/archive rules that prevent the store from becoming a landfill. That conclusion is an inference from the combined reminder, offloading, PIM, organizational-memory, and interruption evidence—not a directly tested branded framework. [3](https://journals.sagepub.com/doi/10.1177/0956797614559285) ## Research method and construct map **Research protocol.** The research question was decomposed by *construct × outcome × time horizon × context*. Primary empirical studies were prioritized over reviews; field and longitudinal evidence was weighted more heavily for professional-transfer claims; preprints were retained only where the area is genuinely emerging, chiefly AI-mediated cognition. Official practitioner sources were used only to establish what PARA/Second Brain claims to be, not as effectiveness evidence. Reviews, testimonials, vendor claims, popularity, Luhmann's personal productivity, and generic RAG/vector-database studies were excluded as proof of human outcomes. The uploaded brief's instruction not to treat related constructs as interchangeable was followed throughout. [1](https://chatgpt.com/g/g-p-6ac129ef43c08191acf988f750b7a6da/c/6ac13845-ef28-83ed-a64f-2ab6d52a698f) The research sequence was: - Define each construct independently and pre-specify outcomes: task success, memory, refinding, learning, decision quality, productivity, or team performance. - Search exact construct names plus causal outcomes, then specific interventions such as reminders, saving, note-taking, folders, tags, search, turnover, decision records, and AI assistance. - Prioritize primary journal articles and original conference studies; use official framework sites only for framework definitions. - Extract population, N, task, intervention, comparator, outcome, duration, publication status, and limitations. - Separate *performance while an aid exists* from *learning or performance after it disappears*. - Separate prospective from retrospective memory, and memory performance from productivity or decision performance. - Investigate contradictions by task, population, time horizon, forced versus voluntary offloading, aid reliability, and note quality rather than counting papers. - Audit professional transfer explicitly: student laboratory evidence receives less weight than naturalistic file retrieval, knowledge-worker teams, federal IT projects, or information-worker field experiments. ``` ``` **Source hierarchy.** For this topic, the preferred source stack is journal/publisher pages and DOI records; PubMed/PsycINFO-style indexing for cognitive psychology; ACM and information-science journals for PIM/HCI; INFORMS and organizational journals for team/organizational evidence; and official working-paper repositories for genuinely new AI-at-work evidence. Preprint servers such as arXiv belong below peer-reviewed research. Practitioner sites belong below all of these for efficacy claims, although an official site is the right primary source for defining its own method. This distinction matters here because Forte Labs explicitly describes Second Brain as an external place to collect and organize ideas and describes CODE as Capture, Organize, Distill, Express; those statements define the intervention family but do not validate it. [7](https://fortelabs.com/) **Construct map.**
Construct What it refers to here What it must not be confused with
**Cognitive offloading** Using action or an external resource to reduce an internal cognitive demand; reminder setting is one instance. Experimental work explicitly operationalizes it this way. [13](https://link.springer.com/article/10.1186/s41235-019-0195-y) Any use of software; learning; shared knowledge
**External memory** An artifact or environmental cue that preserves information or an intention outside biological memory. [3](https://journals.sagepub.com/doi/10.1177/0956797614559285) A guarantee that information will be re-found
**Prospective memory** Remembering that an intended action must be performed later. [2](https://link.springer.com/article/10.3758/s13421-025-01708-x) Remembering what happened in the past
**Retrospective memory** Remembering previously encountered content/events. [12](https://www.psychologicalscience.org/journals/psychological-science/0956797613504438/) Future-action execution
**Note-taking** Both an encoding activity performed now and the creation of an external record for later use. [18](https://www.psychologicalscience.org/journals/psychological-science/0956797614524581/) Retrieval practice
**Retrieval practice** Actively reconstructing knowledge from memory, strengthening later accessibility. [5](https://doi.org/10.1126/science.1152408) Searching for a stored document
**Personal information management** Keeping, structuring, refinding, and using one's files/messages/bookmarks. [6](https://onlinelibrary.wiley.com/doi/abs/10.1002/asi.21415) A particular PKM methodology
**Personal knowledge management** A broader umbrella around personal capture, organization, synthesis and use; operational definitions vary. A single standardized cognitive intervention
**Information foraging** Locating useful information by trading search/navigation effort against expected information value. The PIM findings here primarily operationalize the retrieval side of that problem. [6](https://onlinelibrary.wiley.com/doi/abs/10.1002/asi.21415) Memory strengthening
**Transactive memory** A group system in which people know who has relevant expertise and coordinate access to it. [16](https://pubsonline.informs.org/doi/abs/10.1287/mnsc.1040.0257) A centralized wiki or document store
**Organizational memory** Knowledge retained in people, routines, structures, standardization, relationships, and records. [9](https://academic.oup.com/jpart/advance-article-pdf/doi/10.1093/jopart/muaf019/63489968/muaf019.pdf) Documentation alone
**Knowledge reuse** Applying retained or recovered information to a later task; depending on the study this may appear as transfer, task execution, decision quality, or productivity. [19](https://www.tandfonline.com/doi/abs/10.1207/s15328023top3202_9) Simply possessing an archive
**Distributed cognition** A lens in which cognition can be distributed across people, artifacts, and environment. The empirical studies here test specific mechanisms rather than treating the theory itself as proof of benefit. [3](https://journals.sagepub.com/doi/10.1177/0956797614559285) A claim that every externalization is beneficial
**Second Brain** A practitioner umbrella for capture, organization, distillation and use of digital information. [7](https://fortelabs.com/) A validated psychological construct
**Evidence-extraction template, illustrated with six pivotal sources.**
Source Type Date Key claims Evidence strength Limitations Relevance
Gilbert, *Strategic offloading...* [20](https://pubmed.ncbi.nlm.nih.gov/25404057/) Four experiments 2015 Reminders improve delayed-intention performance; use rises with memory demand/distraction Strong mechanism evidence Artificial online task Core prospective offloading
Storm & Stone [3](https://journals.sagepub.com/doi/10.1177/0956797614559285) Three experiments 2015 Reliably saving earlier information can improve memory for subsequently learned material Strong causal evidence Lab memory task “Free capacity” claim
Urry et al. [21](https://www.psychologicalscience.org/journals/psychological-science/0956797620965541/) Direct replication 2021 Replicated more/verbatim laptop notes, not immediate longhand quiz advantage Strong contradiction/replication Students; no note review Medium-vs-processing debate
Bergman et al. [6](https://onlinelibrary.wiley.com/doi/abs/10.1002/asi.21415) Naturalistic retrieval 2010 Shallow/smaller folders associated with efficient successful navigation High ecological relevance Older desktop environment PIM architecture
Lewis [16](https://pubsonline.informs.org/doi/abs/10.1287/mnsc.1040.0257) Longitudinal team study 2004 TMS development associated with consulting-team performance/viability Good transfer MBA consulting teams, not firms Shared knowledge
Hong [9](https://academic.oup.com/jpart/advance-article-pdf/doi/10.1093/jopart/muaf019/63489968/muaf019.pdf) Organizational panel 2025 Centralization/standardization buffer turnover-related IT delays Strong real-work relevance Observational panel, public sector Organizational memory
The useful visualization depends on the analytical question. **Tables** are superior when comparing study design, population, outcome, or limitations; **charts** are appropriate when studies report comparable quantitative outcomes on a common metric; forcing heterogeneous experiments into a single chart would create false comparability. A **timeline** is useful for seeing how the evidence base expanded from organizational learning and team memory toward digital PIM, cognitive offloading, and AI-mediated work: ``` ``` For scoping future research, it is more defensible to size projects by evidence coverage than promise elapsed durations: a **small** audit might cover 15–25 pivotal primary studies and a claim matrix; a **medium** review 40–80 studies with structured extraction and contradiction analysis; a **large** evidence synthesis would add systematic database searches, duplicate screening, risk-of-bias coding, citation chasing, and potentially quantitative synthesis where outcomes are genuinely commensurable. Those are scope categories, not claims about completion time. ## Individual memory, reminders, notes, and retrieval **Cognitive offloading.** The strongest causal result is that moving part of a memory burden into the environment can improve current performance. Gilbert's delayed-intention experiments showed that external reminders improved prospective-memory performance and became more attractive under higher memory load or anticipated distraction. A newer time-based paradigm required participants to remember actions after 10-, 20-, or 30-second delays while doing a two-back task; reminders improved accuracy and were used increasingly at longer delays. Another experiment involving optional written offloading found particularly strong benefits as short-term memory demand increased. [20](https://pubmed.ncbi.nlm.nih.gov/25404057/) This supports a limited form of the popular statement that “writing things down frees the brain.” The strongest interpretation is not that external storage increases some general-purpose reservoir of intelligence, but that it can **remove a specific competing retention demand**, lowering interference or allowing cognitive resources to be spent elsewhere. Storm and Stone's three experiments are especially informative: saving one file improved learning of a subsequent file only when the save operation was regarded as reliable and the first file otherwise created enough interference to matter. [3](https://journals.sagepub.com/doi/10.1177/0956797614559285) **The reliability condition is critical.** External memory produces value partly because the person expects future access. If that expectation is wrong, the strategy can backfire. In 2026 Fellers and Storm found that reminders improved prospective-memory performance while present, but participants who had previously offloaded an intention subsequently performed worse when those reminders were removed than people who had never offloaded it. The result is unusually relevant to real systems: a tool can simultaneously raise *system-level performance* and reduce *human-only resilience*. [22](https://pubmed.ncbi.nlm.nih.gov/42241083/) **People are not perfectly calibrated about when to offload.** Across reminder paradigms, subjective confidence predicts reminder use, sometimes more clearly than actual unaided memory ability. This means a person can outsource because remembering *feels* difficult even when the externalization is economically or cognitively unnecessary. In a preregistered 2026 study, Ngai and Gilbert found that a very brief prediction-plus-feedback intervention improved calibration and reduced suboptimal reminder-setting; prediction alone was not enough. [13](https://link.springer.com/article/10.1186/s41235-019-0195-y) That weakens another simplistic narrative: “externalize as much as possible.” The empirically defensible objective is **calibrated offloading**, not maximal offloading. An external system is advantageous when its reliability, retrieval cost and failure risk make it better than internal memory for the particular requirement. [23](https://springerlink.fh-diploma.de/article/10.1186/s41235-026-00714-0) **Retrospective memory.** Offloading does not have one uniform effect on what people later remember. Henkel's museum experiments found a photo-taking impairment effect for photographed objects under some conditions: photographing whole objects could reduce later memory for objects, details, and locations compared with observation, while a zoom condition altered that pattern. This supports the possibility that delegating preservation to an artifact changes attention or encoding. [12](https://www.psychologicalscience.org/journals/psychological-science/0956797613504438/) At the same time, offloading can improve memory for *other* material by reducing interference, as Storm and Stone showed. And Silva and colleagues' 2026 48-hour experiment found that alarm reminders and pre-scheduled email improved execution of a future email task without producing a statistically significant difference in later free recall of a news video. Thus “external memory reduces internal memory” is too broad: the effect depends on whether one asks about the offloaded item, subsequent information, contextual content, and whether the aid remains available. [3](https://journals.sagepub.com/doi/10.1177/0956797614559285) **Prospective memory and reminders.** Here the evidence is unusually coherent. Silva et al. randomly assigned 152 participants to an alarm/reminder, scheduled-email, or unaided condition for an email that had to be sent 48 hours later at 7 p.m.; both offloading conditions improved completion. Laboratory work similarly finds reminder benefits for event-based and time-based intentions. This is the strongest empirical foundation for the claim that a task system can prevent commitments from falling through the cracks. [24](https://www.mdpi.com/2076-328X/16/6/872) But the evidence supports **well-cued reminders**, not the entirety of a GTD/Second Brain-style task architecture. A calendar entry, list item, notification, or task record only inherits this evidence insofar as it actually delivers an effective cue at the right moment. There is far less experimental evidence comparing real calendar systems, inboxes, kanban boards, static lists, notifications, and integrated task-management systems against one another on professional outcomes. That is an important transfer gap. **Note-taking.** The original Mueller–Oppenheimer studies found laptop note-takers recorded more information and more verbatim overlap yet did worse on conceptual questions than longhand note-takers, motivating the idea that handwriting forces generative processing. [18](https://www.psychologicalscience.org/journals/psychological-science/0956797614524581/) The replication record makes a medium-centric conclusion unsafe. Urry and colleagues directly replicated the first study with 74 laptop and 68 longhand participants. Laptop users again produced more words and more verbatim overlap, but the longhand group did **not** obtain the original immediate quiz advantage. Their mini-meta-analysis of closely related studies likewise did not support a simple immediate-performance advantage for one medium. [21](https://www.psychologicalscience.org/journals/psychological-science/0956797620965541/) Other work points more directly toward **processing quality**. Katayama and colleagues compared typing study notes with copy-pasting text among 47 undergraduates; after one week, typed construction produced better application-test performance than copy-pasting. Conversely, a 64-student multiple-document study found that note-takers incorporated more source ideas into essays but the no-note group learned more on other measures—another warning that better external capture and better internal learning are different objectives. [19](https://www.tandfonline.com/doi/abs/10.1207/s15328023top3202_9) Structured notes can help under specific educational conditions. Experiments with partial graphic organizers improved exam performance relative to studying complete organizers. But this does not establish that a permanent-note template, progressive summarization, atomic notes, or a project page improves founders' or managers' later decisions. [25](https://www.tandfonline.com/doi/abs/10.3200/JOER.100.2.103-111) **Review matters, but the “weekly review” claim outruns the evidence.** Older educational experiments found benefits from note-taking and later review for delayed recall, but that only establishes that revisiting material can aid learning under some conditions. It does not establish that reviewing one's entire PKM system every seven days is a necessary or optimal productivity intervention. [26](https://bpspsychub.onlinelibrary.wiley.com/doi/abs/10.1111/j.2044-8279.1980.tb00812.x) **Retrieval practice versus storage.** Karpicke and Roediger's experiments show that repeatedly retrieving learned material from memory produces much stronger delayed retention than allocating the same opportunity to repeated study after initial acquisition. The effect is about *successful reconstruction from internal memory*. [5](https://doi.org/10.1126/science.1152408) The practical implication is conditional. Material that must be available **without** tools—core concepts, client context used fluidly in a conversation, domain knowledge needed for reasoning—benefits from internal retrieval practice. Material whose value is primarily in exact preservation—contracts, source citations, historical metrics, verbatim decisions—may be better externalized. A PKM system should not force every datum into biological memory, but neither should external search be described as equivalent to retrieval practice. That distinction follows from the different outcome measured in the retrieval-practice and PIM literatures. [5](https://doi.org/10.1126/science.1152408) ## Personal information architecture, retrieval, and maintenance **The keeping-versus-finding problem.** Personal information management research provides perhaps the most direct warning against “capture everything.” Saving an item and being able to exploit it later are separate stages. In Bergman et al.'s study of 296 people retrieving 1,131 active files, participants succeeded 94% of the time and averaged 14.76 seconds, but performance depended on folder depth and size. Their actual structures were shallow—mean retrieval depth 2.86—and contained relatively small file sets per folder. [6](https://onlinelibrary.wiley.com/doi/abs/10.1002/asi.21415) This is evidence for **retrievability constraints**, not for one universal hierarchy. A useful organizational structure keeps navigational choices manageable and reflects how a person expects to refind things. Very deep taxonomies impose navigation cost; huge flat directories impose selection cost. [6](https://onlinelibrary.wiley.com/doi/abs/10.1002/asi.21415) **Folders versus tags.** In Bergman and colleagues' Gmail study, 75 participants had access to both approaches for a month; in the Windows 7 study, 23 participants were explicitly taught tagging, required to use it for two weeks, and then given five weeks of free choice. Both studies found a strong behavioral preference for folders. When tags were used, people usually used a single tag; multiple classification was rarely exploited in retrieval. In the controlled task, tag-based retrieval was slower and less successful. [14](https://onlinelibrary.wiley.com/doi/10.1002/asi.22906) That does **not** mean tags are universally inferior. It means the oft-repeated assumption that many-to-many classification must outperform hierarchy for human refinding is not empirically established. The theoretical expressiveness of tags is not the same thing as observed user performance. Indeed, a separate survey found people's expressed attitudes toward tags could be more positive than their demonstrated behavior, another reason to prefer observed use over stated preference. [27](https://asistdl.onlinelibrary.wiley.com/doi/10.1002/meet.14505001029) **Search versus navigation.** Earlier PIM research found navigation dominated personal-file retrieval even when stronger search functionality was available: reported navigation shares were roughly 56–68% in the studied environments versus only 4–15% for search, with search often serving as fallback. [28](https://cris.openu.ac.il/ar/publications/improved-search-engines-and-navigation-preference-in-personal-inf-2/) Technology has changed radically since that work, so “folders beat search in 2008” should not be projected mechanically onto semantic search or current AI retrieval. The correct claim is narrower: **historically, merely making search better did not eliminate people's use of spatial/hierarchical organization**, so “search makes organization unnecessary” does not follow from the evidence. A modern replication with semantic/AI retrieval is an obvious research need. [28](https://cris.openu.ac.il/ar/publications/improved-search-engines-and-navigation-preference-in-personal-inf-2/) **Shared-file retrieval makes archive growth costly.** A 2020 naturalistic study had 289 participants retrieve 1,557 of their own shared files. Larger collections, multiple file versions, larger sharing teams, longer time since previous retrieval, folder depth, and greater email workload were all associated with worse retrieval success or efficiency. This is unusually relevant to real team knowledge systems because it studies people's own shared material rather than an artificial information collection. [15](https://www.sciencedirect.com/org/science/article/abs/pii/S2050380620000137) A related bookmark study reinforces the difference between retention and reuse. In a study of 50 participants asked to retrieve previously visited URLs, only a minority of bookmarked targets were actually recovered through the bookmark facility; hierarchical bookmark menus were used particularly rarely. An item being “saved” therefore tells us little about whether it becomes operational later. [29](https://journals.sagepub.com/doi/pdf/10.1177/0961000620949652) **Links and backlinks.** The evidence base drops sharply here. Associative navigation has plausible cognitive and HCI mechanisms, but targeted searches did not locate convincing controlled studies in which backlinking or networked-note navigation was isolated and shown to improve understanding, creativity, professional retrieval, or decision quality relative to folders/search or other sensible alternatives. Evidence about semantic associations, hypertext, or distributed cognition cannot be silently relabeled as evidence that an Obsidian-style backlink graph improves thought. The appropriate classification is therefore **plausible but inadequately tested**, not “disproved.” A clean test would randomly assign knowledge workers doing a months-long research or product task to identical content with either automatic backlinks/network navigation or conventional hierarchy-plus-search, then measure blinded output quality, retrieval latency, source reuse, cross-source synthesis, and maintenance cost. **Zettelkasten evidence audit.** Zettelkasten has a rich historical narrative and modern practitioner literature, but that is not equivalent to intervention evidence. The scholarly material located in targeted searches included educational implementations that bundle Zettelkasten-like methods with other interventions such as interval repetition; such designs cannot isolate the effects of atomicity, permanent-note rewriting, or linking itself. No persuasive independent controlled comparison was located showing that atomic notes improve understanding, that links produce “emergent insight,” that Zettelkasten increases creative output, or that a networked implementation outperforms a hierarchy in professional work. Consequently:
Zettelkasten claim Evidence judgment
Atomic notes improve understanding **Unsupported as a distinctive causal claim**
Rewriting “permanent notes” improves thinking **Plausible but inadequately tested**; generative processing has adjacent evidence, but not the framework intervention itself. [19](https://www.tandfonline.com/doi/abs/10.1207/s15328023top3202_9)
Linking notes generates emergent insight **Plausible but inadequately tested**
Zettelkasten increases creativity/output **Unsupported by controlled framework-level evidence located**
Networked notes outperform hierarchy **Unsupported**; existing PIM evidence does not establish this and some folder/tag findings point the other way. [14](https://onlinelibrary.wiley.com/doi/10.1002/asi.22906)
Luhmann's output validates the method **Invalid causal inference**; one prolific practitioner's biography supplies no counterfactual
**PARA, progressive summarization, and Building a Second Brain.** Forte Labs describes Second Brain as an external place for collecting and organizing important ideas, PARA as Projects/Areas/Resources/Archives, and CODE as Capture/Organize/Distill/Express. The official materials make broad productivity/creative-value claims. [7](https://fortelabs.com/) Targeted searches did not locate an independent randomized trial, quasi-experiment, or credible longitudinal field comparison showing that PARA outperforms topic organization, that project-first organization increases project completion, that progressive summarization improves later retrieval/application, or that the complete Building a Second Brain package increases creativity or knowledge-work productivity. This is an evidence gap, not a finding that the systems cannot work. Some underlying mechanisms **do** have evidence: external reminders help future action; generative processing may aid learning; shallow meaningful structures aid refinding; active retrieval improves internal memory. But stitching several validated mechanisms into a branded workflow does not validate the workflow as a whole. [24](https://www.mdpi.com/2076-328X/16/6/872) **Information overload and maintenance.** Knowledge systems incur costs at capture, categorization, updating, verification and retrieval. The shared-file evidence shows that collection size and multiple versions can directly impair refinding. Field research also demonstrates the cost of turning a knowledge environment into an interruption system: in an experiment with 247 workers, disabling notifications for a workday reduced notification interruptions and improved outcomes related to performance/strain, while a 12-day study of 40 information workers linked more email time to lower perceived productivity and higher measured stress. [15](https://www.sciencedirect.com/org/science/article/abs/pii/S2050380620000137) That makes “maximum capture” a particularly weak design objective. **Expected future reuse value should exceed capture, maintenance, noise, and refinding cost.** The exact optimum has not been experimentally identified for PKM, but the existence of these costs is empirically much better supported than the proposition that ever-larger archives produce ever-larger benefits. [15](https://www.sciencedirect.com/org/science/article/abs/pii/S2050380620000137) **Maintenance and forgetting.** There is no strong evidence that every note requires perpetual preservation or that regular whole-archive review is optimal. Organizational research actually treats some knowledge depreciation and forgetting as normal; Darr, Argote and Epple found substantial depreciation of accumulated know-how in a 1.5-year dataset from 36 service locations. Whether a stale note should be updated, archived or forgotten should therefore be treated as an expected-value problem, not as a moral failure of one's PKM discipline. [30](https://pubsonline.informs.org/doi/abs/10.1287/mnsc.41.11.1750) **Local files versus cloud ownership.** Direct behavioral comparisons of locally owned plain-text archives versus cloud collaborative systems on memory, reasoning, or decision quality were not located. The empirically supported variables are more basic: reliability of future access matters to offloading; refindability and version ambiguity matter in shared collections; and collaboration requires coordination. Claims that local ownership itself improves cognition—or that cloud storage itself improves it—remain inadequately tested. [3](https://journals.sagepub.com/doi/10.1177/0956797614559285) ## Shared knowledge, organizational memory, and decision records **Transactive memory is not “everyone writes everything down.”** Liang, Moreland and Argote compared small groups trained together with individuals trained separately before later group performance on a radio-assembly task. Groups that had learned together developed better transactive-memory structures and performed better. The mechanism is division of cognitive labor coupled with knowledge of expertise location. [8](https://journals.sagepub.com/doi/10.1177/0146167295214009) Lewis then examined 64 MBA consulting teams comprising 261 people longitudinally. Teams with initially distributed expertise and familiar members were more likely to develop transactive memory; its later development depended partly on communication, and stronger TMS was positively related to performance and team viability. This is materially closer to knowledge work than classic memory experiments. [16](https://pubsonline.informs.org/doi/abs/10.1287/mnsc.1040.0257) The implication for a small team's shared system is subtle: a repository is most valuable when it helps answer not only **“where is the document?”** but also **“who understands this?”**, **“who owns it?”**, and **“who should be involved?”** A document store and transactive memory can therefore plausibly complement each other, but the evidence does not show that centralized documentation can replace expertise-location knowledge. [16](https://pubsonline.informs.org/doi/abs/10.1287/mnsc.1040.0257) Research also indicates that transactive-memory benefits are contingent rather than universal. Computational experiments by Ren, Carley and Argote found effects depended on team size, task environment, and whether the objective was speed or quality. The broader lesson is that specialization plus routing can outperform indiscriminate knowledge duplication. [31](https://pubsonline.informs.org/doi/pdf/10.1287/mnsc.1050.0496) **Organizational memory is broader than a wiki.** Darr, Argote and Epple analyzed weekly data over 1.5 years from 36 pizza stores. Unit costs declined with experience, knowledge depreciated over time, and knowledge transfer occurred across stores belonging to the same franchisee but not equivalently across organizational boundaries. That is direct evidence that organizations can accumulate and lose performance-relevant knowledge, though its substrate is not reducible to documents. [30](https://pubsonline.informs.org/doi/abs/10.1287/mnsc.41.11.1750) The professional-transfer case is especially strong in Hong's 2025 federal IT study. Across 94 US agencies and FY2015–2021 projects, centralized CIO authority and task standardization attenuated project delays associated with IT workforce turnover, particularly voluntary and managerial turnover. This indicates that structure and standardized processes can preserve task-relevant organizational memory when individuals leave. [9](https://academic.oup.com/jpart/advance-article-pdf/doi/10.1093/jopart/muaf019/63489968/muaf019.pdf) Other organizational evidence finds abrupt personnel turnover can produce episodic forgetting and performance deterioration, reinforcing that knowledge retained only in people is vulnerable. But this does not establish that every important capability is documentable. The fact that shared work experience and “who knows what” predict team performance indicates that relational and experiential knowledge remains relevant even in organizations with explicit procedures. [32](https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1467-8551.2011.00785.x) The cleanest conclusion about **tacit knowledge** is therefore negative: current evidence does not justify believing that a comprehensive documentation program can convert all valuable know-how into explicit text. Structures and records can preserve repeatable task knowledge; people, shared experience and coordination remain separate performance resources. [16](https://pubsonline.informs.org/doi/abs/10.1287/mnsc.1040.0257) **Decision records and hindsight.** This is one of the better examples of a practitioner idea with a real but narrower empirical mechanism behind it. Davies' three hindsight-bias experiments found that participants who could review records of their pre-outcome cognitions showed less hindsight bias than those without such records. Merely creating a record was not uniformly sufficient; its later availability/use mattered. [10](https://doi.org/10.1016%2F0749-5978%2887%2990005-7) Other work shows why preserving a contemporaneous record could matter. Bradfield and Wells had participants make predictions before learning outcomes, which meant the prediction itself was not subject to memory uncertainty, yet outcome knowledge still changed remembered certainty, reasoning and decision experience. A log containing only the final choice may therefore preserve less than a log recording probabilities, assumptions, evidence, alternatives and rationale. [33](https://pubmed.ncbi.nlm.nih.gov/15915798/) The evidence does **not** establish the stronger claim that decision journals make business decision-makers progressively better or prevent repeated errors. What is established is that outcomes contaminate retrospective reconstruction and that contemporaneous records can partly anchor later evaluation to the actual foresight state. [10](https://doi.org/10.1016%2F0749-5978%2887%2990005-7) For actual practice, the highest-value decision record is therefore probably not a diary entry saying “we decided X.” The mechanism suggests preserving:
Field Why it matters
Decision and date Defines the historical state
Alternatives seriously considered Prevents post-hoc rewriting of the choice set
Key facts believed true Preserves informational context
Assumptions Allows later diagnosis
Probabilities/predictions Makes calibration measurable
Expected upside/downside Separates process from outcome
Reasons/evidence Preserves foresight reasoning
Conditions that would change the decision Creates falsifiable update rules
Owner and review trigger Connects memory to action
This template is an evidence-informed inference from hindsight-bias research rather than a directly validated “decision journal” protocol. [10](https://doi.org/10.1016%2F0749-5978%2887%2990005-7) ## AI-mediated external memory and transfer to professional work AI changes three parts of the external-memory loop simultaneously: it can reduce the cost of **capture/summarization**, reduce the cost of **retrieval**, and generate a new interpretation of the stored material. The first two are potentially valuable; the third means retrieval is no longer necessarily faithful reproduction. The strongest professional AI studies are adjacent rather than specific to PKM. A field experiment involving 758 BCG consultants found large productivity and quality gains on tasks that fell within GPT-4's capability frontier, alongside worse performance on a task outside that frontier. This establishes that AI can augment knowledge-work execution, not that an AI-indexed personal archive necessarily improves memory or decision quality. [34](https://aiinstitute.hbs.edu/navigating-the-jagged-technological-frontier/) A later large field experiment at Procter & Gamble studied professionals doing product-development work and found that AI could alter collaboration and expertise-sharing patterns; the study is directly relevant to distributed expertise, but again it tests generative collaboration rather than long-term external-memory retrieval. [35](https://aiinstitute.hbs.edu/the-cybernetic-teammate-how-ai-is-reshaping-collaboration-and-expertise-in-the-workplace/) **Reduced effort is not automatically better cognition.** In a small 2025 preprint experiment on AI-supported lecture note-taking, 30 participants experienced varying degrees of assistance. Fully automated structured notes were preferred for ease and reduced effort, but the intermediate-assistance condition produced the strongest post-test outcome, while the highest-assistance condition performed worst. Because the study is small, educational and a preprint, it should be treated as frontier evidence rather than settled fact—but it directly illustrates the possibility that optimization for friction reduction conflicts with encoding. [17](https://arxiv.org/abs/2509.03392) Related preregistered preprint research with 184 participants found that AI involvement in idea-generation workflows reduced later source-memory accuracy, particularly when human and AI contributions were intermixed. Again, the population/task is not a longitudinal company knowledge base, but it raises a relevant provenance problem: AI can make knowledge easier to produce while making authorship and source origin harder to remember. [36](https://arxiv.org/abs/2509.11851) **Verification remains necessary.** Glickman and Sharot's series of experiments with 1,401 participants demonstrated feedback loops in which interaction with biased AI shifted and amplified subsequent human perceptual, emotional, and social judgments. The important PKM implication is not that every AI summary is biased in this way; it is that repeated AI mediation can alter the user, so a system cannot be evaluated solely by whether retrieval feels easier. [37](https://www.nature.com/articles/s41562-024-02077-2) For AI summarization itself, hallucination and factual-fidelity research continues to show that generated summaries require accuracy evaluation. More importantly for this report, convincing longitudinal human experiments showing that automatic meeting summaries improve later organizational decisions while preserving correct nuance and attribution remain sparse. [38](https://www.nature.com/articles/s41598-025-31075-1) **Professional transfer matrix.**
Finding Original context Transfer to founder/manager/small team
Reminders increase prospective-memory success Mostly laboratory tasks; one 48-hour email experiment **Good mechanism transfer.** Forgotten commitments are structurally similar, though real tools add notification overload. [24](https://www.mdpi.com/2076-328X/16/6/872)
Saving reduces interference Laboratory learning **Moderate.** Plausible for multitasking/research; no direct productivity estimate. [3](https://journals.sagepub.com/doi/10.1177/0956797614559285)
Offloading can impair unaided later memory Laboratory memory **Moderate.** Relevant when tool outages or internal expertise matter. [22](https://pubmed.ncbi.nlm.nih.gov/42241083/)
Longhand > laptop Student lectures **Weak.** Replication does not justify workplace handwriting mandates. [21](https://www.psychologicalscience.org/journals/psychological-science/0956797620965541/)
Generative processing > verbatim capture Educational note tasks **Moderate mechanism, weak workplace estimate.** [19](https://www.tandfonline.com/doi/abs/10.1207/s15328023top3202_9)
Retrieval practice strengthens internal memory Educational/lab learning **Good for internalized expertise; weak for PKM search.** [5](https://doi.org/10.1126/science.1152408)
Shallow folders facilitate refinding People's own files **High**, though technologies studied predate modern semantic search. [6](https://onlinelibrary.wiley.com/doi/abs/10.1002/asi.21415)
Collection/version/team size impair shared-file retrieval People's own shared files **High.** Directly resembles real work repositories. [15](https://www.sciencedirect.com/org/science/article/abs/pii/S2050380620000137)
TMS relates to knowledge-team performance MBA consulting teams plus lab teams **Moderate–high.** Directly relevant to small expert teams. [16](https://pubsonline.informs.org/doi/abs/10.1287/mnsc.1040.0257)
Organizational structure buffers turnover US federal IT organizations **High mechanism transfer**, exact effect sizes organization-dependent. [9](https://academic.oup.com/jpart/advance-article-pdf/doi/10.1093/jopart/muaf019/63489968/muaf019.pdf)
Foresight records reduce hindsight reconstruction Experimental judgment **Moderate.** Mechanism strong; repeated business-decision benefit untested. [10](https://doi.org/10.1016%2F0749-5978%2887%2990005-7)
AI boosts some professional task performance Consultants/product professionals **High for bounded AI-enabled tasks**, not yet evidence for personal archives. [34](https://aiinstitute.hbs.edu/navigating-the-jagged-technological-frontier/)
AI assistance can reduce learning/source memory Students/online experimental participants; some preprints **Low–moderate, provisional.** [17](https://arxiv.org/abs/2509.03392)
The biggest transfer problem in the entire literature is that **most cognitive experiments optimize one outcome at a time**, whereas professional systems must jointly optimize execution, learning, resilience, collaboration, verification, retrieval speed and maintenance cost. A reminder can improve execution while weakening learning; exhaustive documentation can improve continuity while increasing retrieval noise; AI can reduce search cost while increasing verification demands. Treating any single laboratory benefit as sufficient to recommend an entire PKM architecture is therefore unjustified. [22](https://pubmed.ncbi.nlm.nih.gov/42241083/) ## Claim audit and material-study table **Popular Second Brain claims.**
Claim Classification Evidence-based judgment
**“Writing things down frees the brain for higher-level thinking.”** **Conditionally supported** Reliable offloading can reduce interference/cognitive demand and improve concurrent or subsequent performance, especially under higher load. This does not imply a generic increase in “higher-level thinking.” [3](https://journals.sagepub.com/doi/10.1177/0956797614559285)
**“People should capture everything.”** **Contradicted as a general prescription** Collection size, version proliferation, maintenance and retrieval noise carry costs; saved information is frequently not reused. [15](https://www.sciencedirect.com/org/science/article/abs/pii/S2050380620000137)
**“Everything important should live in one system.”** **Plausible but inadequately tested** Consolidation may reduce fragmentation, but no direct comparison establishes one-system superiority across real work.
**“Projects are a better organizing principle than topics.”** **Unsupported** No independent controlled PARA test located. Forte defines PARA's project orientation, but definition is not validation. [39](https://help.fortelabs.com/hc/en-us/articles/34874403416461-What-is-the-Second-Brain-Notion-Template-and-what-s-included)
**“Backlinks improve thinking.”** **Plausible but inadequately tested** No convincing direct evidence located that backlinking improves understanding, creativity or decisions relative to alternatives.
**“Atomic notes improve understanding.”** **Unsupported** Adjacent evidence supports generative processing, not atomicity as the causal variable. [19](https://www.tandfonline.com/doi/abs/10.1207/s15328023top3202_9)
**“External memory reduces cognitive load.”** **Directly/conditionally supported** Supported when a task genuinely contains a memory burden and the aid is usable/reliable; the benefit disappears or reverses under some conditions. [3](https://journals.sagepub.com/doi/10.1177/0956797614559285)
**“A Second Brain improves creativity.”** **Plausible but inadequately tested** No independent framework-level causal test located. [7](https://fortelabs.com/)
**“Decision logs reduce repeated mistakes.”** **Plausible but inadequately tested** Contemporaneous records can reduce hindsight reconstruction; long-term error reduction is not established. [10](https://doi.org/10.1016%2F0749-5978%2887%2990005-7)
**“Weekly reviews are necessary.”** **Unsupported as a necessity claim** Reviewing can aid memory, but no evidence establishes a seven-day whole-system review as necessary or optimal. [26](https://bpspsychub.onlinelibrary.wiley.com/doi/abs/10.1111/j.2044-8279.1980.tb00812.x)
**“Structured metadata improves retrieval.”** **Conditionally supported / mixed** Structure affects retrieval, but more metadata is not automatically better; tags were slower/less successful in one controlled PIM comparison. [14](https://onlinelibrary.wiley.com/doi/10.1002/asi.22906)
**“Search makes folders unnecessary.”** **Contradicted in the tested systems** Historical studies found navigation strongly preferred even with improved search. Modern semantic/AI search requires fresh testing. [28](https://cris.openu.ac.il/ar/publications/improved-search-engines-and-navigation-preference-in-personal-inf-2/)
**“Storing more knowledge creates more value.”** **Contradicted as a monotonic claim** More stored material can increase retrieval costs; storage alone is not reuse. [15](https://www.sciencedirect.com/org/science/article/abs/pii/S2050380620000137)
**“AI makes a large archive operationally useful.”** **Plausible but inadequately tested** AI can improve bounded knowledge-work tasks, but longitudinal archive-to-decision evidence is absent and AI mediation creates verification/bias risks. [34](https://aiinstitute.hbs.edu/navigating-the-jagged-technological-frontier/)
**Material-study extraction table.** Where the surfaced primary abstract did not provide a defensible sample size, the table says so rather than reconstructing or manufacturing one.
Study / link Status / date Design, population, N Task / intervention / comparator Outcome / duration Result Key limitation / professional transfer
Gilbert, “Strategic offloading of delayed intentions...” [20](https://pubmed.ncbi.nlm.nih.gov/25404057/) `https://doi.org/10.1080/17470218.2014.972963` Peer-reviewed, 2015 4 online experiments; total N=1,196 Delayed-intention task; external reminders vs internal memory; varying load/distraction Prospective-memory performance; immediate, with one naturalistic follow-up component Reminders improved performance; offloading rose with load/distraction Artificial task. **Transfer: moderate**
Henkel, “Point-and-Shoot Memories” [12](https://www.psychologicalscience.org/journals/psychological-science/0956797613504438/) `https://doi.org/10.1177/0956797613504438` Peer-reviewed, 2014 Museum experiments; N not extracted from publisher abstract Photograph whole objects / zoom / observe Later object/detail/location memory Whole-object photography impaired some later memory; zoom altered effect Photography ≠ generic note storage. **Transfer: low–moderate**
Storm & Stone, “Saving-Enhanced Memory” [3](https://journals.sagepub.com/doi/10.1177/0956797614559285) `https://doi.org/10.1177/0956797614559285` Peer-reviewed, 2015 3 experiments; N not in publisher abstract Save earlier file before studying new file; manipulate reliability/interference Memory for subsequent material; immediate Saving improved new learning when reliable and interference was sufficient Lab lists/files. **Transfer: moderate**
Boldt & Gilbert, “Confidence guides spontaneous cognitive offloading” [13](https://link.springer.com/article/10.1186/s41235-019-0195-y) `https://doi.org/10.1186/s41235-019-0195-y` Peer-reviewed OA, 2019 Experimental; N in full article, not extracted here Instructed vs spontaneous reminder use PM performance and offloading choice Both strategies aided performance; confidence guided offloading Artificial task. **Transfer: moderate**
Tsai & Gilbert, time-based intentions [2](https://link.springer.com/article/10.3758/s13421-025-01708-x) `https://doi.org/10.3758/s13421-025-01708-x` Peer-reviewed, 2025 Experiments; N not extracted here 10/20/30-s intention while performing two-back; optional reminders Time-based PM Reminders improved accuracy; longer delay increased offloading; cue visibility mattered Seconds-long delays. **Transfer: moderate mechanism**
Silva et al., alarms vs scheduled email [24](https://www.mdpi.com/2076-328X/16/6/872) `https://doi.org/10.3390/bs16060872` Peer-reviewed OA, 2026 Randomized, N=152 Email exactly 48h later; reminder vs scheduled email vs unaided PM completion + retrospective video recall, 48h Both offloading groups improved prospective performance; no significant retrospective-memory difference Student sample; single commitment. **Transfer: moderate–high**
Fellers & Storm, reminder removal [22](https://pubmed.ncbi.nlm.nih.gov/42241083/) `https://doi.org/10.1037/xlm0001630` Peer-reviewed, 2026 2 experiments; N not recovered from surfaced abstract External reminders, followed by removal Current PM + later unaided performance Reminders helped while present; prior offloading could impair subsequent unaided performance Laboratory task. **Transfer: moderate, strategically important**
Ngai & Gilbert, metacognitive training [23](https://springerlink.fh-diploma.de/article/10.1186/s41235-026-00714-0) `https://doi.org/10.1186/s41235-026-00714-0` Peer-reviewed OA, 2026; preregistered Exp. 1 N=164; Exp. 2 N=416 Prediction/feedback interventions before reward-based reminder decisions Calibration and optimal reminder choice Prediction + feedback improved calibration and absolute offloading bias; prediction alone insufficient Incentivized lab paradigm. **Transfer: moderate**
Mueller & Oppenheimer, laptop vs longhand [18](https://www.psychologicalscience.org/journals/psychological-science/0956797614524581/) `https://doi.org/10.1177/0956797614524581` Peer-reviewed, 2014 3 student studies; N not extracted here Longhand vs laptop lecture notes Factual/conceptual test; mostly immediate Laptop notes more verbatim; original paper found poorer conceptual performance Replication qualification; students. **Transfer: low**
Urry et al., direct replication [21](https://www.psychologicalscience.org/journals/psychological-science/0956797620965541/) `https://doi.org/10.1177/0956797620965541` Peer-reviewed, 2021 Direct replication, N=142 (74 laptop; 68 longhand), plus mini-meta-analysis Same core laptop/longhand comparison Immediate quiz More words/verbatim typing replicated; longhand performance advantage did not No later study/review condition. **Transfer: low–moderate**
Katayama et al., electronic note construction [19](https://www.tandfonline.com/doi/abs/10.1207/s15328023top3202_9) `https://doi.org/10.1207/s15328023top3202_9` Peer-reviewed, 2005 N=47 undergraduates Type notes vs copy-paste; note scaffolds Facts/application after one week Typing notes improved application transfer vs copy-paste Small educational sample. **Transfer: moderate mechanism**
Karpicke & Roediger, retrieval practice [5](https://doi.org/10.1126/science.1152408) `https://doi.org/10.1126/science.1152408` Peer-reviewed, 2008 Controlled learning experiments; N not extracted here Repeated retrieval vs repeated study Delayed recall Repeated retrieval produced superior long-term retention Educational word learning; not document search. **Transfer: moderate for internal expertise**
Bergman et al., folder structure [6](https://onlinelibrary.wiley.com/doi/abs/10.1002/asi.21415) `https://doi.org/10.1002/asi.21415` Peer-reviewed, 2010 Naturalistic; N=296; 1,131 files, 5,035 navigation steps Retrieval of participants' own active files Success/time during real-file retrieval 94% success, mean 14.76 s; depth and folder size affected performance Older desktop technology. **Transfer: high for basic PIM**
Bergman et al., folders vs tags [14](https://onlinelibrary.wiley.com/doi/10.1002/asi.22906) `https://doi.org/10.1002/asi.22906` Peer-reviewed, 2013 Gmail N=75; Windows N=23 Folder/tag use observed; Windows tagging training and controlled retrieval Storage choice, retrieval speed/success over weeks Strong folder preference; tags slower and less successful in controlled task Small Windows sample; older systems. **Transfer: moderate–high**
“Factors hindering shared files retrieval” [15](https://www.sciencedirect.com/org/science/article/abs/pii/S2050380620000137) `https://doi.org/10.1108/AJIM-05-2019-0120` Peer-reviewed, 2020 Naturalistic, N=289; 1,557 own shared files Retrieve actual shared files Success, time, missteps Collection size, versions, sharing-team size, recency, depth, workload hindered retrieval Correlational within naturalistic task. **Transfer: high**
Liang, Moreland & Argote, group training/TMS [8](https://journals.sagepub.com/doi/10.1177/0146167295214009) `https://doi.org/10.1177/0146167295214009` Peer-reviewed, 1995 Experimental small groups; N not extracted here Train together vs separately; later radio assembly Recall and group performance after delay Joint training improved TMS and later group performance Manual task, small groups. **Transfer: moderate**
Lewis, knowledge-worker TMS [16](https://pubsonline.informs.org/doi/abs/10.1287/mnsc.1040.0257) `https://doi.org/10.1287/mnsc.1040.0257` Peer-reviewed, 2004 Longitudinal; 64 MBA consulting teams, N=261 Observe TMS formation across consulting project Performance and viability across project Developed TMS positively related to performance and viability Student consulting teams. **Transfer: moderate–high**
Darr, Argote & Epple, organizational learning [30](https://pubsonline.informs.org/doi/abs/10.1287/mnsc.41.11.1750) `https://doi.org/10.1287/mnsc.41.11.1750` Peer-reviewed, 1995 36 pizza stores, weekly data over 1.5 years Learning-by-doing across stores/franchisees Unit production cost over time Learning occurred, knowledge depreciated, transfer followed organizational boundaries Operational service work rather than knowledge work. **Transfer: moderate**
Hong, federal IT turnover [9](https://academic.oup.com/jpart/advance-article-pdf/doi/10.1093/jopart/muaf019/63489968/muaf019.pdf) `https://doi.org/10.1093/jopart/muaf019` Peer-reviewed, 2025 Panel of IT projects in 94 US federal agencies, FY2015–2021 Variation in turnover, centralization, standardization Project delays over years Structures retaining task knowledge buffered turnover-related delays Non-randomized public-sector study. **Transfer: high mechanism**
Davies, hindsight/foresight records [10](https://doi.org/10.1016%2F0749-5978%2887%2990005-7) `https://doi.org/10.1016/0749-5978(87)90005-7` Peer-reviewed, 1987 Three experiments; N not recovered from surfaced abstract Pre-outcome cognition record / review vs controls Hindsight bias Reviewing foresight records reduced bias; merely recording was not sufficient in all conditions Laboratory judgment. **Transfer: moderate**
Ohly & Bastin, notification interruption [40](https://onlinelibrary.wiley.com/doi/full/10.1002/1348-9585.12408) `https://doi.org/10.1002/1348-9585.12408` Peer-reviewed, 2023 Field experiment N=247 Notifications disabled for a workday vs usual condition Interruptions, performance/strain Removing notification interruptions improved relevant work outcomes; effects varied with individual factors One-day intervention. **Transfer: high**
Email-use field study [41](https://doi.org/10.1145/2858036.2858262) `https://doi.org/10.1145/2858036.2858262` Peer-reviewed CHI, 2016 40 information workers over 12 workdays Natural email behavior; batching patterns Perceived productivity, physiological stress More email time associated with lower productivity perceptions/higher stress; batching associated with higher perceived productivity Observational. **Transfer: high context, lower causal certainty**
Chen et al., AI-assisted note-taking [17](https://arxiv.org/abs/2509.03392) `https://arxiv.org/abs/2509.03392` **Preprint**, 2025 Within-subject study N=30 Automated AI notes vs intermediate summary vs minimal assistance Cognitive engagement and post-test Most automated aid easiest/preferred but worst post-test; intermediate assistance best Tiny educational sample, non-peer-reviewed. **Transfer: low/provisional**
Glickman & Sharot, human–AI feedback [37](https://www.nature.com/articles/s41562-024-02077-2) `https://doi.org/10.1038/s41562-024-02077-2` Peer-reviewed, published online 2024 / volume 2025 Series of experiments, total N=1,401 Repeated human interaction with AI judgments Human perceptual/emotional/social judgments AI interaction could amplify human biases through feedback loops Not a PKM/archive study. **Transfer: moderate for AI-reliance risk**
The table itself illustrates a key methodological asymmetry: the **strongest evidence for external-memory mechanisms comes from controlled cognitive tasks, while the strongest evidence for real work comes from PIM, team, organizational and interruption studies**. Very few studies combine both high causal identification and months-long professional use. ## What current evidence does not support and the research frontier **Contradictory evidence is mostly conditional rather than mysterious.** Several apparent contradictions disappear when the dependent variable is made explicit:
Apparent contradiction Likely resolution supported by the studies
Offloading helps memory / offloading hurts memory It often improves **system-level task success** while potentially reducing **internal encoding of the offloaded information**. [3](https://journals.sagepub.com/doi/10.1177/0956797614559285)
Saving causes forgetting / saving improves learning Memory for the saved material and memory for subsequently learned material are different outcomes; reducing interference can help the latter. [3](https://journals.sagepub.com/doi/10.1177/0956797614559285)
Handwriting beats typing / it does not The original conceptual advantage did not replicate cleanly; generative versus verbatim processing is a better candidate mechanism than medium itself. [4](https://www.psychologicalscience.org/journals/psychological-science/0956797614524581/)
Tags are more flexible / folders work better Many-to-many categorization is structurally expressive, but human retrieval behavior may favor remembered locations and simple hierarchies. [14](https://onlinelibrary.wiley.com/doi/10.1002/asi.22906)
Search should eliminate folders / users still navigate Retrieval is often an orienteering process based on contextual/location cues; historical improved search did not remove that preference. [28](https://cris.openu.ac.il/ar/publications/improved-search-engines-and-navigation-preference-in-personal-inf-2/)
Documentation preserves knowledge / turnover still hurts Some task knowledge can be embedded in structure and records, while expertise, relationships and coordination remain person-dependent. [9](https://academic.oup.com/jpart/advance-article-pdf/doi/10.1093/jopart/muaf019/63489968/muaf019.pdf)
AI lowers cognitive burden / AI can harm cognition Lower effort can increase immediate efficiency while reducing encoding, provenance awareness or independent checking. [17](https://arxiv.org/abs/2509.03392)
**Current evidence does not support** treating personal knowledge management as an accumulation contest. It does not establish that every useful thought should be captured, that one universal repository is better than several specialized systems, that projects are always a superior organizing dimension, that folders are obsolete, that backlinks are cognition-enhancing, that “atomicity” has a measurable independent benefit, that progressive summarization is optimal, or that a weekly review cadence is universally necessary. Direct framework-level evidence for these propositions is absent or inadequate. Nor does the evidence establish that an excellent external-memory system should maximize internal forgetting. Fellers and Storm's reminder-removal result points in the opposite direction for knowledge and intentions that must survive tool failure. A robust professional design needs to distinguish **things the human must know** from **things the human merely needs to be able to recover**. [22](https://pubmed.ncbi.nlm.nih.gov/42241083/) **The most important research gaps and studies that would resolve them are:**
Gap Study that would meaningfully resolve it
**Does a complete Second Brain/PARA system improve professional outcomes?** Multisite preregistered 6–12 month randomized or cluster-randomized trial among knowledge workers comparing PARA/BASB training with a minimal capture/search system and business-as-usual. Measure missed commitments, retrieval latency, project cycle time, output quality, perceived load and maintenance time.
**Do backlinks/network graphs improve understanding or creativity?** Randomize the *same corpus and search capability* to hierarchy-only versus hierarchy+backlinks/graph. Run months-long synthesis projects and blind-rate final insight novelty, source integration, refinding, and time cost.
**Do atomic notes add value independently of rewriting?** Factorial experiment crossing atomic vs multi-idea notes with verbatim vs generative rewriting, followed by delayed synthesis and transfer tasks.
**Are projects superior to topics for organization?** Longitudinal within-person crossover with identical information assigned to project-first versus topic-first organization, measuring retrieval time and project completion—not satisfaction alone.
**What is the optimal capture threshold?** Manipulate capture selectivity under controlled workload: exhaustive, moderate/value-filtered, and minimal capture. Measure later useful retrievals against capture/maintenance/search cost.
**How should internal memory and external storage be allocated?** Train participants under varying aid reliability and then unexpectedly remove the aid. Measure immediate performance, delayed internal knowledge, recovery time, and decision quality. The Fellers–Storm paradigm provides a starting mechanism. [22](https://pubmed.ncbi.nlm.nih.gov/42241083/)
**Does a decision journal improve future decisions?** Multi-year prospective field trial with decision-makers randomly assigned to structured predictions+rationale+scheduled review versus outcome-only logging or normal practice; score calibration, process quality and recurrence of diagnosed error types.
**Can documentation substitute for departing experts?** Study matched team transitions with documentation-only handoff, expert-overlap handoff, both, and standard practice; measure ramp-up, errors, time-to-autonomy and escalation frequency.
**How should centralized docs interact with transactive memory?** Factorial team study varying repository quality and explicit expertise-directory quality. Test whether they substitute, complement or interfere.
**Does modern semantic/AI search eliminate hierarchy benefits?** Replicate classic PIM navigation/search studies using current semantic and conversational retrieval on participants' real multi-year corpora. [6](https://onlinelibrary.wiley.com/doi/abs/10.1002/asi.21415)
**Does AI-assisted archive retrieval improve decisions?** Longitudinal professional experiment comparing keyword search, semantic search and AI answer-with-citations on the same company corpus. Measure not only answer speed but correctness, verification, source inspection, decision quality and false-confidence rate.
**Do automatic meeting summaries improve organizational memory?** Randomize teams to raw transcript, human notes, AI summary, and AI-summary-plus-verification. After weeks/months test factual reconstruction, responsibility attribution, decision rationale and action completion.
**What review cadence is optimal?** Compare event-triggered review, weekly review, adaptive review based on reuse probability, and no systematic review; measure benefits minus maintenance time rather than memory alone.
The **research frontier** is therefore not another note taxonomy. The highest-value unanswered question is how to optimize the *whole human–external-memory system* across four competing objectives: ## [ \\text{Net value} \\approx \\text{action reliability} + \\text{useful retrieval/reuse} + \\text{decision quality} + \\text{organizational continuity} ## \\text{capture cost} ## \\text{maintenance cost} ## \\text{retrieval noise} \\text{dependency/error cost}. ] That equation is a conceptual synthesis rather than an estimated empirical model, but each positive and negative term corresponds to effects observed in the research reviewed above. [24](https://www.mdpi.com/2076-328X/16/6/872) For a founder, manager, or small team, the evidence therefore points to a **lean external-memory architecture rather than an encyclopedic Second Brain**: **Use hard external cues for commitments that must happen.** Reminder evidence is strong. [2](https://link.springer.com/article/10.3758/s13421-025-01708-x) **Store exact information that is expensive or dangerous to reconstruct—sources, decisions, assumptions, agreements, metrics—but preserve provenance.** Hindsight research and AI-mediated judgment make provenance unusually important. [10](https://doi.org/10.1016%2F0749-5978%2887%2990005-7) **Internalize knowledge that must be available fluently without retrieval.** Retrieval practice, rather than passive note accumulation, is the better-supported mechanism for that objective. [5](https://doi.org/10.1126/science.1152408) **Optimize for refinding, not filing sophistication.** Shallow structures, manageable collections, sensible visibility, version control and clear naming have more direct evidence behind them than elaborate metadata or link graphs. [6](https://onlinelibrary.wiley.com/doi/abs/10.1002/asi.21415) **For teams, preserve both documents and expertise location.** A shared repository without “who knows what” leaves out a demonstrated component of team cognition. [16](https://pubsonline.informs.org/doi/abs/10.1287/mnsc.1040.0257) **Archive and delete aggressively enough that retrieval remains useful.** There is no empirical virtue in a large archive merely being large; growth creates measurable retrieval costs. [15](https://www.sciencedirect.com/org/science/article/abs/pii/S2050380620000137) **Use AI as a retrieval and synthesis layer with source visibility, not as an unquestioned replacement for the underlying record or for independent judgment.** Professional productivity benefits are real on suitable tasks, but human–AI feedback and emerging encoding evidence make maximal automation a poor default. [11](https://aiinstitute.hbs.edu/navigating-the-jagged-technological-frontier/) The evidence therefore supports a rather less glamorous principle than “build a Second Brain”: **externalize selectively, make future use cheap and reliable, preserve what the artifact cannot preserve by itself—context, provenance, expertise, and judgment—and evaluate the system by successful later action, not by how much knowledge it contains.**