22  From Notes to Research Infrastructure

22.1 A Library Begins to Remember

At 9:20 on the morning of November 29, 2022, a cluster of items entered my Readwise library. One promised “networked thought.” Another explained how smart notes could turn reading into writing. Others addressed exports, associative links, and the Zettelkasten. Their shared premise was simple: the researcher needed an external memory that preserved ideas beyond the moment of reading.¹

The tools changed. The underlying problem did not. A social scientist may spend years gathering articles, archival records, interview transcripts, observations, code, and partial arguments. The problem is not only how to store these materials. It is how to recover the right piece at the right time, reconstruct why it mattered, connect it to a changing question, and preserve the line from source to claim.

AI entered the same library within days. The early items treated it as a knowledge-aware writing assistant and an unusually productive partner for idea generation. The promise was speed and variation. The warning was already present: an output could be inventive and wrong at the same time.²

I argue that PKM and AI solve different research problems. Personal knowledge management supplies persistent context, provenance, and continuity. AI reduces the cost of searching, comparing, transforming, and testing that context. Each remains incomplete without the other. PKM without computation can become an orderly archive that rarely produces an argument. AI without a governed knowledge base can produce fluent answers that have no stable relation to the research record.

Their integration creates something more useful than a better note-taking system or a more informed chatbot. It creates research infrastructure. The infrastructure remembers what the researcher has read. It records what the researcher decided. It lets machines perform bounded work across those records. It also preserves the human gates that separate a candidate from evidence and evidence from a claim.

This chapter develops that argument from an audit of my Readwise library. On July 27, 2026, the library contained 567 entries tagged 📄Academic Article, 499 tagged 🗃️PKM, and 3,196 tagged 🤖AI. These labels overlap, and the AI category contains news, product announcements, social-media threads, and duplicate records as well as scholarly material. The counts therefore describe attention, not evidence. I use the complete PKM set, the connector-accessible academic and AI samples, their overlaps, and sixteen inspected records to reconstruct the movement from capture to research infrastructure.³

22.2 From Capture to Connection

The first generation of PKM tools treated forgetting as the central constraint. Researchers read more than they could remember, saved more than they could retrieve, and divided related ideas among folders created at different moments. The proposed remedy was to externalize memory. Notes would become smaller, searchable, and linked. Daily writing would reduce the cost of capture. Backlinks would allow an idea recorded for one purpose to reappear in another.

This was a major improvement over a hierarchy of project folders. A folder asks where a note belongs. A network asks what the note is related to. That difference matters in social research because the same source can perform several jobs. An interview may provide evidence about an event, reveal an actor’s vocabulary, expose a retrospective justification, and suggest a new comparison. A single folder cannot express all four relations.

The early library also connected note-making to writing. The point was not to collect quotations. It was to leave one’s future self a usable intellectual object: a claim, a question, a disagreement, a definition, or an unresolved connection. One item from December 2022 defined PKM as a response to information abundance and unreliable memory. Its stronger implication was that thinking had to become visible outside the head before it could become cumulative.⁴

For academic work, this first phase contributed four capacities. It preserved memory across semesters and projects. It let concepts travel between literatures. It separated source material from the researcher’s interpretation. It made intellectual development observable through dated notes and links.

These capacities support more than productivity. They support reflexivity. A qualitative researcher can return to an early memo and see which explanation once seemed plausible, which case disrupted it, and when the coding scheme changed. A historical researcher can distinguish a contemporaneous reading from an interpretation formed after later evidence appeared. A collaborative team can trace a concept to the person, source, and decision that introduced it.

But connection did not solve selection. A graph could display thousands of notes without explaining which ones bore on a claim. A search could find every mention of “violence” without distinguishing an event report from a theoretical definition. PKM made the research memory durable. It did not yet make that memory easy to interrogate.

22.3 When AI Entered the Notebook

The next phase reduced the cost of transforming source material. OCR converted scanned pages into text. Speech recognition converted interviews and lectures into transcripts. Translation brought sources in additional languages into a shared workspace. Summarization produced initial descriptions. Semantic search retrieved passages that did not share the researcher’s exact vocabulary.

The library recorded both the gain and the danger. In December 2023, an item described using a multimodal model to transcribe a handwritten journal. The system outperformed conventional text recognition because it used context to infer unclear words. It also changed the date from 20 to 28. Context rescued words whose meaning could be inferred. Context corrupted a number that could not.⁵

This is the central methodological problem of AI-assisted transformation. The same capacity that repairs incomplete text can invent plausible detail. A model may standardize an institutional name correctly in ninety-nine records and silently merge two different organizations in the hundredth. It may translate a political term fluently while erasing the ambiguity that made the term analytically important. It may summarize an interview and omit the hesitation that changes how the answer should be read.

The solution is not to reject transformation. It is to preserve the original, the transformed object, the procedure, and the verification status as separate artifacts. The scan remains the source. The OCR text remains a derivative. A correction log records changes. A human validates fields whose error would alter the analysis.

The same period changed scholarly reading. Entries in the library moved from global search and backlinks toward semantic readers, citation maps, paper explainers, and interfaces that connected articles to videos and related work. The PDF remained the publication object, but it no longer had to remain the research interface.⁶

For social scientists, this shift expanded access to evidence whose form had once made it prohibitively expensive. A researcher can now search a large newspaper archive, transcribe oral histories, extract tables from reports, compare terminology across languages, and locate names in thousands of pages. A recent library entry describing the digitization of nearly twenty million historical census images shows the scale of this change. It also shows why the work remains a research design problem: models, prompts, schemas, validation samples, and error rates shape the resulting data.⁷

AI made more material computable. It did not make all computed material evidence.

22.4 From Retrieval to Workflow

By 2025, the library’s focus had shifted from individual features to academic workflows. One entry paired a citation manager with a local Markdown vault. The citation manager held bibliographic authority and source files. The vault held reading notes, concepts, project memos, and draft arguments. The value came from the division of labor, not from forcing one application to perform every function.⁸

Another entry placed an AI agent inside a structured note environment and asked it to find weaknesses in arguments drawn from writing, meetings, and journals. This was a different relationship between researcher and machine. The model no longer responded only to a prompt. It operated over a body of prior material and returned an analytical artifact for inspection.⁹

The change can be stated in three sentences. First, PKM externalized memory. Next, AI made that memory searchable and transformable. Then workflows connected repeated operations to named inputs, outputs, and checks.

This is where the integration becomes valuable for academic research. A prompt is disposable. A workflow can be inspected, repeated, compared, and revised. A prompt asks for sources. A workflow receives a review protocol, maps vocabulary, searches approved databases, records every query, stages candidates, resolves identities, tests coverage, and stops before final inclusion. A prompt asks for themes. A workflow preserves the transcript, coding instructions, proposed codes, rejected codes, uncertainty, and researcher approval.

The distinction also changes how researchers evaluate AI. The unit of analysis is not the model alone. It is the system consisting of the model, context, tools, permissions, stopping rule, output schema, verifier, and human gate. A weaker model inside a well-specified workflow may produce more trustworthy research artifacts than a stronger model inside an unconstrained conversation.

22.5 Persistent Context and the Agentic Turn

The 2026 PKM records in the library differ sharply from the early collection. Among the 79 PKM entries saved that year, 48 also carry the AI tag, 56 refer to Obsidian, and 14 concern agents. The object of interest is no longer the isolated note. It is the relationship between a durable, local knowledge base and an agent that can act across it.

Several records describe the vault as a permanent memory layer. A project file states the question, scope, standards, active decisions, and prohibited actions. Markdown notes preserve the source record. An agent reads the smallest set of files required for a task. Repeated procedures become skills. Scheduled workflows process an inbox, prepare a briefing, identify connections, or stage a research memo.¹⁰

One of the strongest entries in this group proposes a continuous system in which Readwise supplies captured material, Obsidian preserves it, an AI layer reasons across it, and an automation layer moves artifacts between stages. My note on that item identified the deeper claim: knowledge gains value over time when the cost of maintenance falls and new work can repeatedly encounter old material.¹¹

The attractive phrase is “a second brain that thinks back.” The phrase is useful, but it can mislead. A vault does not become reliable because an agent can read it. Old notes may contain errors. Duplicates may look like corroboration. A summary may have lost the source’s qualification. A private record may be outside the agent’s authorized scope. Persistent context amplifies whatever the system contains.

The operative design therefore requires three components. The first is a specification that defines acceptable work before execution. The second is a knowledge base that supplies bounded context and authoritative artifacts. The third is a verifier that tests the output against the specification. A recent library item formulated this pattern as “spec, verifier, knowledge base.” It also recorded a decisive warning: a verifier measures only what it was designed to measure.¹²

This pattern now appears in more ambitious claims about AI scientists. One saved item distinguishes retrieval, search within a fixed schema, and discovery that changes the schema itself. Its proposal remains provisional for the social sciences, but the distinction is productive. Retrieval finds a known object. Search explores a defined space. Discovery proposes that the space was defined incorrectly. Only the third requires the researcher to reconsider the concepts, boundaries, or variables that organized the project.¹³

AI may propose such a revision. It may not ratify it.

22.6 What PKM Contributes to Social Science

Personal knowledge management contributes temporal depth. Social-science projects often outlast software, funding cycles, and individual bursts of attention. A durable note records not only a result but the state of the inquiry when the result appeared. This matters when a concept changes meaning or when later events make an earlier interpretation look inevitable.

PKM contributes conceptual continuity. Political and social concepts travel under different names across disciplines, countries, institutions, and periods. Linked notes can connect an indigenous term, an archival label, a theoretical category, and a coding variable without pretending they are identical.

PKM contributes provenance. A useful research note identifies its source, locator, date of capture, transformation history, and relation to the project. The note does not replace the source. It points back to it.

PKM contributes negative evidence. Researchers usually save what supports an argument. A governed system can also preserve failed searches, rejected explanations, excluded cases, unresolved contradictions, and missing records. This negative evidence protects against repeating work and against rewriting the research history after the conclusion becomes clear.

The same principle applies at institutional scale. A retrospective account of a long-running conflict-research program attributes its achievements not to data volume alone but to the combination of social-scientific method, practitioner knowledge, flexible support, and contextual interpretation. Data became useful because an institutional memory connected collection to the circumstances that gave the records meaning.¹⁴

PKM contributes cumulative comparison. A case memo can share fields with other case memos while retaining material that does not fit the common schema. This combination is important in qualitative and comparative work. Structure supports comparison. Free text preserves surprise.

Finally, PKM contributes ownership and durability when it relies on open, inspectable formats. Plain-text notes can survive the application that created them. They can be versioned, searched, diffed, backed up, and read by both humans and machines. This is not a preference for one note application. It is a preservation decision.

22.7 What AI Contributes

AI contributes scale. It can prepare candidate records from more documents than one researcher can read line by line. It can locate names, dates, events, institutions, quotations, and relations across a large corpus. Candidate is the operative word. Extraction creates material for validation.

AI contributes flexible retrieval. Keyword search depends on shared vocabulary. Language models can suggest synonyms, historical terms, translations, rival concepts, and proxy measures. This capacity can prevent a false gap produced by a poor query. It can also drift beyond the review’s scope, which is why vocabulary expansion must remain logged and reviewable.

AI contributes comparison. A model can apply the same questions to a set of documents, identify points of agreement and conflict, and prepare a matrix that a researcher can inspect. It can ask which source supports a date, which cases violate an emerging pattern, or which paragraphs depend on one unverified claim.

AI contributes transformation. OCR, transcription, translation, classification, entity resolution, and structured extraction can move evidence among forms. The result becomes methodologically useful only when the transformation retains a link to the original.

AI contributes iteration. It can rerun a procedure after the codebook changes, compare versions, and report which records changed. This makes dynamic coding possible without hiding the consequences of revision.

AI also contributes disciplined opposition when instructed to do so. It can search for disconfirming evidence, propose rival mechanisms, detect scope changes, and ask which observation would distinguish two explanations. It cannot decide the substantive importance of those findings. But it can make it harder for a researcher to avoid them.

22.8 The PKM-AI Research Loop

The integrated system works when every stage produces an artifact and every consequential transition includes a human gate. The following PKM-AI research loop can support a literature review, archival project, interview study, event database, or comparative case analysis.

Stage	Durable artifact	Permitted AI work	Human gate
Bound	Versioned project-context packet	Test ambiguity; propose missing terms and boundaries	Approve the question, scope, ethics, and stopping rule
Capture	Source note with stable identity and locator	Import metadata; transcribe or OCR into a derivative file	Confirm rights, identity, and source-derivative relation
Distill	Claim, evidence, question, and uncertainty notes	Propose summaries, entities, claims, and contradictions	Compare each consequential claim with the source
Connect	Concept map and evidence map	Suggest links, aliases, rival concepts, and missing source families	Accept relations and reject false equivalence
Compute	Candidate register, query log, or coding table	Search, classify, compare, extract, and propose new variables	Approve inclusion, exclusions, codebook changes, and promotions
Verify	Validation sample, discrepancy log, and coverage memo	Run schema checks, known-item tests, consistency checks, and adversarial review	Judge adequacy, residual error, and whether to stop
Express	Versioned manuscript, dataset, appendix, or public release	Draft from approved artifacts; generate alternative explanations and formats	Take responsibility for claims, citations, interpretation, and release

The loop is recursive. A contradiction found during writing may change the concept map. A new archival term may change the search protocol. A validation failure may require a new codebook version. The system should permit revision while preserving the prior state.

Consider a researcher building an event database from daily newspapers. The PKM layer holds the research question, source inventory, codebook versions, decisions, and notes about access or political context. The AI layer scans permitted newspapers, stages event candidates, extracts dates and actors, proposes additional variables, and identifies possible duplicates. The researcher validates a sample, resolves ambiguous events, approves codebook changes, and records why a candidate entered or left the verified corpus.

The result is not merely a larger dataset. It is a dataset with a memory.

22.9 An Operative Chapter Exercise

This chapter can be tested on one active research question. Begin with no more than twenty source notes. Write a one-page context packet containing the question, population, period, concepts, source boundaries, ethics, current artifact versions, and decision authority. Select five known items that any adequate search must recover.

Ask an AI system to map the vocabulary and stage a candidate register. Require a stable source identifier, discovery route, bibliographic status, relevance rationale, uncertainty, and verification status for every row. Do not permit a final inclusion field.

Inspect the result in three passes. First, check identity and provenance. Second, look for missing terms, false equivalences, and scope drift. Third, examine what the system failed to recover. Record every correction in an audit log and rerun the same task with the revised context packet.

The exercise succeeds if the second run becomes easier to inspect and if the record explains why it changed. It does not succeed because the prose sounds better or because the model returns more candidates.

The downloadable Skills and Agents Lab provides a literature-discovery skill, bounded scout agent, project-context template, three synthetic benchmarks, and a failure report for this exercise.

22.10 What AI Must Not Do

AI must not convert a retrieval failure into a claim that no literature exists. It must not treat repeated records as independent corroboration. It must not overwrite the canonical note, codebook, or corpus without a versioned proposal. It must not use restricted or personal material merely because that material is technically accessible. It must not hide uncertainty behind a completed field.

AI must not collapse source, derivative, note, and claim into one object. It must not turn a retrospective interview into a contemporaneous record. It must not translate a disputed concept into a stable English equivalent without preserving the original. It must not infer causal direction from narrative order alone.

Most of all, AI must not become the unnamed author of consequential research decisions. A final inclusion, exclusion, codebook revision, causal judgment, or public claim requires a responsible researcher. The system may prepare the decision. It may preserve the reasons. It may expose inconsistency. The researcher still decides.

22.11 The Research System That Remembers

The evolution recorded in my library begins with a familiar anxiety: too much to read, too much to remember, and too many tools. It ends, for now, with a different question. How should a researcher govern a system that can read and act across years of accumulated intellectual work?

The answer is not a fully autonomous scholar. It is a partnership built from asymmetry. PKM remembers slowly and precisely. AI processes quickly and probabilistically. PKM preserves the path. AI explores possible paths. PKM holds the source record. AI prepares transformations and comparisons. Human judgment defines the question, interprets the evidence, and accepts responsibility for the claim.

This integration pushes social research forward by changing the feasible scale of careful work. It lets one researcher inspect more archives, follow more conceptual variants, compare more cases, and rerun more coding decisions. But the gain comes from governed iteration, not automation alone.

The durable principle is simple. Build a knowledge base that another person can inspect. Give AI a bounded task within that base. Require it to return an artifact, not an answer. Verify the artifact against the source and the specification. Then write the next version without erasing the last.

---

1. Roam Research, “A Note-Taking Tool for Networked Thought,” accessed July 27, 2026, https://roamresearch.com; Tiago Forte, “How To Take Smart Notes: 10 Principles to Revolutionize Your Note-Taking and Writing,” February 4, 2020, https://fortelabs.co/blog/how-to-take-smart-notes/.

2. Mem, “Introducing the World’s First Knowledge-Aware AI Assistant,” December 8, 2022, saved as a forwarded email in Readwise Reader; Ethan Mollick, “How to Use AI to Generate Ideas,” December 3, 2022, https://oneusefulthing.substack.com/p/how-to-use-ai-to-generate-ideas.

3. Readwise Reader library audit conducted July 27, 2026. Reader reported 567 records tagged 📄Academic Article, 499 tagged 🗃️PKM, and 3,196 tagged 🤖AI. Detailed metadata retrieval returned all 499 PKM records and connector-limited samples of 500 academic and 500 AI records. The audit treated tags as organizational labels rather than quality judgments, inspected sixteen representative full-text records, and made no changes to the library. See editorial/readwise-pkm-ai-source-register.md.

4. Sébastien Dubois, “What Is Personal Knowledge Management (PKM)?,” December 16, 2022, https://pkmjournal.com/why-is-personal-knowledge-management-pkm-useful-5f23405c2a2f.

5. Tiago Forte, “Great Use Case for ChatGPT Vision: Transcribing Handwritten Journal Entries,” December 11, 2023, https://x.com/fortelabs/status/1734284384537333813.

6. Andy Stapleton, “Is This the End of Traditional Academic Papers? See What Every Researcher Needs to Know!,” video, December 6, 2023, https://www.youtube.com/watch?v=kQp7EbxHcKA.

7. Noah Dasanaike, “What I’ve Learned From Digitizing 20 Million Historical Documents,” February 4, 2026, https://noahdasanaike.github.io/posts/digitizing-census.html.

8. Len_dde, “My Academic Workflow with Obsidian and Zotero,” April 5, 2025, https://medium.com/%40lennart.dde/my-academic-workflow-with-obsidian-and-zotero-53ac44ebdc28.

9. Andrew Altshuler, “The Argument Analyst Tana Template Is Out,” March 7, 2025, https://x.com/1eo/status/1898034427579531367/.

10. CyrilXBT, “Most People Use Obsidian as a Note-Taking App,” May 5, 2026, https://x.com/cyrilXBT/status/2051589237880009026/. The library note marks this item as a redundant, shorter version of a longer guide; the chapter uses it only to document the persistent-context argument.

11. CyrilXBT, “Obsidian Plus Vellum: A Second Brain That Never Stops,” May 27, 2026, https://x.com/cyrilxbt/status/2059486502401569047/.

12. “Karpaty’s Second Brain,” shared Claude conversation saved July 18, 2026, https://claude.ai/share/9263ec7c-827e-48da-8ddf-65375205f96b. This is a derivative conversation about a video, not an independently verified account of the original presentation.

13. Markus J. Buehler, “We’ve Made a Breakthrough in Self-Evolving AI Scientists,” June 5, 2026, https://x.com/profbuehlermit/status/2062865983459475830/. The saved item cites F. Y. Wang and M. J. Buehler, “Self-Revising Discovery Systems for Science: A Categorical Framework for Agentic Artificial Intelligence,” arXiv:2606.01444 (2026). The claim requires verification against the paper before scholarly citation.

14. Eli Berman and Jacob N. Shapiro, “Big Data Seeks Context, for Long-Term Relationship: Reflections on the Empirical Studies of Conflict Project (ESOC),” August 17, 2023, https://www.stimson.org/2023/reflections-on-the-empirical-studies-of-conflict-project-esoc/. The byline and date were verified against the publisher page; the record has not yet been checked against Zotero.