Track C / 4 volumes

Build your own second brain

A track for founders and leads who want a personal knowledge system they can actually ask. A pile of exports is not a second brain. A second brain is a protected, versioned source store with retention controls, a compiled wiki that interprets them, a lint that keeps the wiki honest, and a retrieval path that returns evidence before a model speaks. The tools that pull mail or notes will change. This pipeline should not.

Sample available Linked samples are ready to read. Other modules show planned outcomes and are marked “Coming soon”.

Volume 1

From sources to a wiki you can ask

Sample lesson available · remaining lessons coming soon

Complete the minimum preflight before importing any source. Use fictional notes first; authorise each source, its readers and retention. Immutable means protected from casual agent edits, not exempt from controlled correction, deletion or backup expiry.

The durable idea: collect once, compile with provenance, lint what you compiled, and retrieve a small piece of evidence — do not let an agent wander the source store.

C1.01Read the lesson

From sources to a wiki you can ask

Completed artifact: Source-to-answer sketch with authorised readers. Outcome: You can sketch the path from a source, through an immutable store and a compiled wiki, to a linted page an agent can retrieve.

  • The wiki is the thing you ask; source versions remain subject to authorised access, correction and retention.
  • If the agent has to open everything to answer, you do not yet have a second brain.
C1.02Coming soon

Immutable sources

Completed artifact: Source register with retention and controlled correction procedure. Outcome: You can explain why source exports are kept as a record, and why an agent should not freely rewrite or browse them.

  • Collecting is cheap; changing the past is expensive.
  • If a source is sensitive, the right move may be to leave it out.
C1.03Coming soon

The compiled wiki

Completed artifact: One compiled page with three supported claims. Outcome: You can name the kinds of pages a useful wiki needs — an index, sources, entities, concepts, analyses, a journal, short memories, and a log.

  • The wiki is interpretation, with pointers back to the sources.
  • An index entry is not the same thing as a page that has been read and filed.
C1.04Coming soon

Ingest, compile, and lint

Completed artifact: Scoped lint report including one deliberately broken link. Outcome: You can describe ingest, compilation, and a health check as three different jobs with three different owners.

  • Ingest brings sources in; compilation writes the interpretation; lint looks for drift, dead links, and orphans.
  • A lint that reports a clean bill on the wrong scope is worse than no lint.
  • Give each stage an exit check. "Collected" means the source and its owner are known. "Compiled" means every claim points back to its source. "Healthy" means stale pages and contradictions have been flagged.
C1.05Coming soon

Provenance, contradiction, and the log

Completed artifact: Conflict note with evidence pointers and owner. Outcome: You can keep a pointer from a claim to its source, and record a contradiction instead of silently picking a side.

  • When two sources disagree, the page should say so.
  • A short activity log is how the next session knows what was already tried.
C1.06Coming soon

The human curates; the agent files

Completed artifact: Human-review record accepting or refusing a proposed page. Outcome: You can decide which judgements stay with you, and which filing work an agent may do.

  • An agent can draft a page and propose links; you accept the interpretation.
  • Credential vaults, identity documents, and anything you would not show a colleague stay out of the store.
C1.07Coming soon

Retrieval after the wiki

Completed artifact: Three retrieval questions with support or explicit abstention. Outcome: You can test a bounded retrieval path that scores an authorised index and attaches evidence, expanding to multiple sources when the question requires it.

  • The model enters after the evidence is in hand.
  • A pointer hop is useful; an open-ended search through sources is a last resort.
C1.08Coming soon

Your first source-to-wiki path

Completed artifact: Complete fictional source, index, page, lint and answer bundle. Outcome: You can take one kind of source — a weekly note, a meeting summary, or a mail label — all the way through the pipeline.

  • Prefer one boring, complete path over five half-connected exports.
  • Stop when you can ask a real question and get a cited answer.

Volume 2

Privacy, secrets, and holds

Coming soon

A personal vault is still a store of other people's data. This volume is about what never goes in, what is held back, and how a secret scan becomes a gate rather than a suggestion.

C2.01Coming soon

What must never be ingested

Outcome: You can name classes of material that stay out of the vault entirely.

  • Credentials, identity documents, and unrelated family or staff files are not "context."
  • If you would not put it in a shared drive, do not put it in a second brain.
C2.02Coming soon

A secret scan as a real gate

Outcome: You can put a credential scan in the path of every commit, with a failing exit code that actually stops the commit.

  • A scan you can walk around is not a gate.
  • The scan should report a path and a category, never the secret itself.
C2.03Coming soon

Scan what will be committed, not what you hope is there

Outcome: You can explain why the scan must see the staged bytes, and why a piped or empty run is not a pass.

  • Cleaning the working copy after staging is not the same as cleaning what will be committed.
  • Add, then scan, then commit — as one chain. Do not scan a truncated log and call it green.
C2.04Coming soon

Redact at export, not only at commit

Outcome: You can prefer redaction at the point a source is copied, so a later gate is a backstop, not the first defence.

  • Support exports and mail often contain passwords in the body of a ticket.
  • A hold, a skip, or a redacted snapshot is better than hoping the scan will catch it.
C2.05Coming soon

Privacy holds

Outcome: You can pause a class of exports until a redaction-safe refresh is reviewed, and keep a hold list that does not itself leak.

  • Hashed identifiers plus a review date are enough to manage a hold.
  • A hold is a decision with an owner, not a forgotten folder.
C2.06Coming soon

The human enters passwords and completes 2FA

Outcome: You can keep sign-in in the person's hands, including when an agent is doing the filing.

  • Agents do not copy credentials, and do not use a switch that skips permission prompts.
  • If a tool asks the person to sign in again, that is the path — not a relocated secret file.
C2.07Coming soon

Personal data and PDPA in a personal vault

Outcome: You can treat a founder's vault under the same minimisation habits you would ask of a company system.

  • Singapore's PDPA, and Thailand's PDPA when people in Thailand are involved, still apply to other people's data you hold.
  • This is a briefing for practitioners, not legal advice.
C2.08Coming soon

A hold is sometimes the right decision

Outcome: You can choose not to ingest a source, and write that decision down, without treating it as a failure of the pipeline.

  • Completeness is not the goal; a safe, useful wiki is.
  • Review holds on a date, rather than leaving them to fade from view.

Volume 3

The health of the wiki

Coming soon

Collecting will always outrun filing. This volume is about measuring the right gap, and not trusting a number from a check you have not scoped.

C3.01Coming soon

Collecting is easy; filing is the work

Outcome: You can tell a full source store from a wiki that has actually been read and interpreted.

  • An unread export is not organisational memory.
  • The backlog that matters is usually filing, not missing files.
C3.02Coming soon

Export gaps and filing gaps are different jobs

Outcome: You can separate a pipeline defect from an analyst backlog, and pick the fix that matches the cause.

  • An export gap means the source never arrived. A filing gap means it arrived and nobody filed it.
  • Measure before you build a new job to nag yourself.
C3.03Coming soon

A stale date is a hypothesis

Outcome: You can measure a page against its live sources before you rewrite it.

  • A date that is months behind is a reason to check, not a reason to start over.
  • Some pages only need a number and a date refreshed; some need a real re-read.
C3.04Coming soon

Volume is not activity

Outcome: You can decompose a jump in volume before you treat it as a finding.

  • Automated mail and recovered history will inflate a count.
  • Ask what the denominator is made of before you report growth.
C3.05Coming soon

Check the checker

Outcome: You can ask what a health check cannot see, and test it against a case you know should fail.

  • A check that under-reports is indistinguishable from success.
  • A clean result on a new check is a reason for suspicion, not celebration.
C3.06Coming soon

State what a check covers

Outcome: You can report a count together with its scope, and make exclusions explicit and counted.

  • "Backlog cleared" is only true inside the corpora you actually checked.
  • Silent exclusions are how a precise number becomes a wrong one.
C3.07Coming soon

File timestamps lie when exporters rewrite

Outcome: You can avoid treating an exporter's rewrite as new information.

  • Many exporters rewrite a file in full on every run, so the file's timestamp tracks the run, not the content.
  • Look at when the content last changed, then look at the diff, before you call a page stale.
C3.08Coming soon

A short health review

Outcome: You can run a short, regular review: lint, filing gaps, holds, and a few stale pages worth measuring.

  • Keep the output small enough to act on.
  • Do not bulk-update "last reviewed" dates for work you did not do.

Volume 4

Asking the brain

Coming soon

The point of the pipeline is an answer you can trust, with evidence attached, at a cost you can defend.

C4.01Coming soon

Deterministic retrieval before the model

Outcome: You can put a no-model retrieval step in front of any answer, so the model is working from evidence.

  • Keywords, an index, and a scoring pass are enough to start.
  • The model should not be the thing that hunts through the vault.
C4.02Coming soon

Start with a bounded evidence budget

Outcome: You can test one best page as an initial budget, then expand to multiple authorised sources when needed for support or contradiction checks.

  • Opening everything is slow, expensive, and hard to audit.
  • Measure coverage and support; a cross-source question may need several sections.
C4.03Coming soon

Evidence, then an answer

Outcome: You can attach an evidence block to the model's prompt, and refuse to answer from memory when the block is empty.

  • The evidence is what you can show a colleague.
  • If retrieval missed, say so — do not pad.
C4.04Coming soon

Benchmarking the path

Outcome: You can compare a retrieval path with free exploration on the same questions, the same model, and the same harness.

  • Record quality, time, tokens, and tool calls, not only a pass mark.
  • Illustrative only: a single small internal practitioner run, not a performance guarantee. The same model and harness answered eight real questions from a personal vault: one arm explored freely, while the other scored an index, opened one best file and attached an evidence block. The free-exploration arm already had a compiled wiki and the questions were not held out. Both arms scored 8/8; reported average times were 63.6 versus 25.2 seconds (about 2.5× faster for retrieval), variable tokens were 96 thousand versus 66 thousand (about 31 percent fewer), and tool calls were 22 versus 8. The underlying evidence has not been independently verified, and the source store was not fully searched. These results do not establish causation or predict quality, cost or speed on other work.
C4.05Coming soon

Honest caveats

Outcome: You can write down what a benchmark does not prove.

  • If the free-explore arm already had a compiled wiki, say so.
  • If the questions were not held out, say so. A useful result is still useful when its limits are visible.
C4.06Coming soon

When retrieval is the wrong tool

Outcome: You can name questions that need a different path — a new source, a full-text search, or a person.

  • An index of titles and summaries will miss a fact that lives only in an uncited export.
  • Knowing when to stop asking the wiki is part of operating it.
C4.07Coming soon

Writing a memory the next session can use

Outcome: You can record an operating lesson as a short, dated memory, indexed and ready to retrieve.

  • A memory is a lesson, not a diary dump.
  • One command that writes the page, the index line, and the log is better than a ritual you will skip.
C4.08Coming soon

Tools will change; the pipeline should not

Outcome: You can keep sources, wiki, lint, and retrieval intact when an exporter or a model is replaced.

  • The pipeline is the durable asset.
  • A new tool has to earn its place in ingest, not in a side folder nobody lints.