AI Second Brain

A personal knowledge store an agent can read — in the simplest and most durable form, just markdown files and folders, organized so that both you and your agent can find things again. The definition is deliberately unglamorous: it is not a product or a vector database, it is plain files, which is exactly why it is tool-agnostic — the same store works across Claude Code, Codex, and other agent harnesses because "it's just files and folders." The single test of whether you have one is retrieval: can your agent find it again, and could you? If the answer is no, the fix is usually Context Routing (the agent was never told where to look), not more machinery. A recurring, easily-misattributed truth: the bottleneck is normally getting knowledge out of your head into the system, not the system's retrieval — before blaming the AI, check whether your files actually hold the nuance your brain does.

How sophisticated the store needs to be is a separate axis, captured by Retrieval Maturity Levels; how you should shape what you store is Query-Shaped Storage; what you should let in is Evergreen vs Volatile Context. A second brain is the routing intelligence over your knowledge, not merely the ingested copy of it.

Claims

A second brain is just files and folders organized so you and your agent can find things again — nothing more exotic is required. principle — durable: the value is organization and retrievability, not the storage technology; because it's plain markdown it is portable across any agent.
The only real test is "can it find it again?" principle — durable: retrievability is the definition; a store you can't recall from is not a second brain regardless of how it looks.
The bottleneck is usually getting knowledge out of your head into the system, not the system's retrieval. principle — durable and counter to the common instinct to blame the model; the missing nuance is often absent from the files, not lost in search.
Keep the second brain tool-agnostic — it's just files, so it survives switching agent harnesses. best practice — context: operators who move between Claude Code, Codex, and others; the "best" store is the portable one precisely when you don't want lock-in. See Context Routing for the CLAUDE.md→AGENTS.md→memory.md mechanics.
Extraction can be a skill, not manual note-taking — interview yourself to fill the store. best practice — context: when the constraint is offloading tacit knowledge; Nate uses a "Grill Me" skill that interrogates him until it "knows everything" and writes the file.
"Just files and folders" is the right default for a human+agent store; it is contested for a purely agent-to-agent store, where a second source argues structured data beats markdown. best practice — context: the reader determines the shape. Nate's second brain is read by both human and agent, so portable markdown wins; Jaymin West's Context Substrate is agent-first and machine-queried, where he argues markdown rots and structured records win. Same discipline, opposite call, split by who reads it — hold both, don't collapse to one rule.
Split the store into verbatim capture and synthesized knowledge with hard rules — raw/ = capture verbatim, never reorganize; wiki/ = AI-written only, never edit by hand, synthesized and source-cited; outputs/ = disposable, generated on demand. best practice — context: a second, independent source (Austin Marchese) reaches the same raw-vs-wiki separation Nate uses, and enforces it via a CLAUDE.md of rules so the agent maintains the discipline. Contingent on a human+agent store where provenance (source-cited wiki) matters.
Extraction can be a skill: record yourself talking, then let the agent interview you to fill the gaps before ingesting. best practice — context: a second source (Marchese) corroborates the "getting knowledge out of your head is the bottleneck" claim with the same interview move as Nate's "Grill Me" skill — record a life-story/goals monologue, then "analyze this and interview me to fill in anything I missed."
The raw/ (verbatim capture) vs wiki/ (AI-synthesized) split, plus an index.md and an append-only log.md, is now a third-independently-arrived-at layout — and it traces to Karpathy's published "LLM wiki" gist. best practice — context: a human+agent store the agent incrementally maintains; Nate Herk's walkthrough reaches the same raw/+wiki/+index+log structure Marchese uses, citing Karpathy's gist as the origin. The index and log are the agent's maintenance surface — it crawls them plus backlinks to find and update data. When the wiki is compiled (not just filed), this becomes a Compiled Knowledge Base.
"It's just markdown files with routing," so the brain is portable to any agent, not only the one that built it. principle — durable and independently restated: because the store is plain .md, an operator can point Codex or another agent at the same files; the value lives in the files and their routing, not the harness.
A second brain that is reliably available and consulted as a matter of course is, on the extended-mind view, part of the operator's cognition — not merely a tool. principle — as asserted by Andy Clark's Parity Principle (judge by role, not mechanism): Patrick Jones's Evernote web is a second brain avant la lettre, and damaging such a store is damage to the extended person+store system, not a tooling inconvenience. Grounds why markdown-as-truth is load-bearing. See Extended Mind / Cognitive Scaffolding.
"Your mind is for having ideas, not holding them." principle — David Allen, quoted by the 2023 Contrary Research report: working memory is small and long-term memory decays, so holding belongs in the external store. The offload premise beneath the whole cluster, stated pre-agent-era.
Manual PKM fails on alignment, labor intensity, and fragmentation, decaying into learn→write→forget; the fix is Dynamic Retrieval, not more capture. observation — the 2023 Contrary Research report's failure analysis of static stores (Notion/Roam/Obsidian done by hand); the pre-agent statement of the problem this pattern solves. See Capture-Storage-Retrieval Pipeline for the pipeline decomposition.
Value grows when the LLM is treated as a thinking engine over the store — suggesting perspectives, finding gaps, connecting unrelated pieces — rather than a text generator, and compounds as the store grows. principle — as asserted by the same report; the 2023 anticipation of what the agent-era sources describe building.
A second brain differs from a wiki by where knowledge sits on a stabilization gradient: "a wiki is where knowledge goes when it stops changing quickly, whereas a second brain is where it still changes shape." principle — as asserted by the glukhov.org article: wikis favor explicit structure, shared naming, and convergence toward a source of truth, which makes them awkward for half-formed, private, exploratory thinking; a second brain begins personal, evolving, and ambiguity-tolerant, existing before consensus settles.
Second brain, wiki, and RAG are complementary, not interchangeable: a second brain can feed a wiki, a wiki can supply a clean source for RAG, and RAG can make a second brain easier to search — while RAG alone retrieves at inference time and "does not preserve human interpretation by itself." principle — as asserted by the same article; the second brain remembers what mattered, why it mattered, and how interpretation shifted, which query-time retrieval structurally does not. See Compiled Knowledge Base for the compile-at-ingest resolution of the same gap.
Over time value shifts from individual notes to the relationships among them — the store acts like a personal knowledge graph without requiring a literal graph view. observation — the same article's evolution account: notes become connections, connections become reusable patterns, patterns cultivate judgment. An independent restatement of the compounding claim, and of this cluster's "the link structure is the graph" posture.
The raw/+wiki/+index+log split (traced to Karpathy) plus the "grill me" interview extraction is now a fourth-independently-arrived-at layout — Taoufik rebuilds the same store in Obsidian. best practice — context: a human+agent store the agent grows incrementally; Taoufik's on-screen vault shows raw/ (the inbox "you write here") and wiki/{concepts, sources, entities, templates, …} with an index and log, credits "the idea: Andrej Karpathy," and uses a "grill me" skill (run first) that interrogates the operator to build the profile. Same layout, same origin story, same extraction move as Nate Herk / Marchese / the Fable-5 gist — convergence, not a new claim.
Bigger graph ≠ better: past a point, more nodes add noise that hurts the agent's retrieval, not signal. observation — Taoufik opens on the "the bigger, the better?" instinct and crosses it out: "if the AI agent can't actually use the right information or there's too much noise, it will be very hard to track." A restatement, at the store level, of Evergreen vs Volatile Context — curate what gets in; density is not the goal.
Stage ingestion behind explicit operator approval — watch → report → approve → ingest → verify — so nothing enters the durable store until the operator agrees, and dedup against the wiki happens at watch time. best practice — context: agent-driven ingestion of external sources (here, videos) into a human+agent store where provenance and trust matter; Taoufik's "watch" skill drops a report.md into raw/ ("not yet processed, and it won't process before I'm agreeing"), tells the operator which pages it would add for verification, and reports "already documented in your wiki" instead of duplicating. Human-in-the-loop staging is the guard against silent, unreviewed growth. See Multimodal Video Ingestion.
A typed template per capture kind (article, asset, book, brainstorm, …) lets the agent fill a fixed structure instead of inventing one per note. best practice — context: an Obsidian store using the Templater plugin; before writing, the agent classifies the input's type and instantiates that type's template, so structure is consistent across the store. The note-type analogue of adapt-a-template-don't-invent (cf. Golden Templates for the code-exemplar form).
Guard the store's structural health with a lint gate before writes, not only its content. best practice — context: a link-structured store where orphans break retrieval; Taoufik runs a deterministic "brain-lint" pre-write hook that checks the brain is "healthy enough to add" and flags orphan pages. Spun out as Knowledge-Graph Lint.

Compiled Knowledge Base — the eager-synthesis ingest contract (read→extract→integrate) layered on top of the "just files" store; the LLM-wiki pattern is a second brain that compiles.
Knowledge-Graph Lint — the structural-health gate that keeps this store's link graph connected (orphan check) before the agent writes.
Distillate: I Gave Claude Code a Permanent Memory — a fourth independent rebuild of the raw/wiki/index/log + Karpathy + grill-me layout; adds the approval-staged ingest pipeline and the brain-lint gate.
Retrieval Maturity Levels — how sophisticated the brain's retrieval needs to be; the ladder this store climbs (or, usually, shouldn't).
Context Routing — the mechanism that makes a file-and-folder brain retrievable: a CLAUDE.md router that says where things live.
Query-Shaped Storage — the principle for shaping what goes in: store backwards from how you'll ask.
Evergreen vs Volatile Context — the discipline for what to let in vs leave accessible.
Decision Log — a first-class artifact inside the brain: append dated decisions on every big change.
Agentic Simplicity — "boring markdown is beautiful" is the simplicity thesis applied to personal knowledge.
Context Substrate — the agent-first, structured relative of the second brain; the two hold the graph's live tension over structured-vs-markdown, split by who reads the store.
Distillate: Fable 5 + Karpathy's LLM Wiki is Basically Cheating — a third source for the raw/wiki/index/log split and the "just markdown files with routing → portable" claim.
Distillate: Every Level of a Claude Second Brain Explained
Distillate: This Claude Skill Watches Videos So You Don't Have To — where "second brain" was first held as a theme before this page spun it out.
Distillate: The Agentic Engineering Meta — the contrarian case for an agent-first store: structured data over markdown when only agents read it.
Distillate: How to Build a Self-Improving System with Claude Code — a second source for the raw/wiki split and the record-yourself-then-interview-me extraction move.
Self-Improving System — the second brain wrapped in continuous ingestion pipelines and a self-improvement loop; a second brain is the store at its centre.
Distillate: Hermes Architecture EXPLAINED: Memory, Context & Gateways — another independently-built corroborating instance: Hermes' soul.md/user.md/memory.md are agent-maintained markdown memory, always in context (see Layered Agent Memory).
Extended Mind — the principle altitude above this pattern: an external, reliably-consulted store counts as part of the cognitive system by role, not mechanism.
Distillate: Andy Clark — What is Extended Mind? (Closer To Truth) — the philosophical lineage (Clark's parity principle and the ethical status of external memory), pre-dating the LLM tooling by decades.
Capture-Storage-Retrieval Pipeline — the pipeline skeleton this store fleshes out; localizes where manual PKM breaks.
Dynamic Retrieval — the retrieval bar: surfaced by task and context, not by the operator's recall. Holds a dated tension with this page: the 2023 report assumed dynamic retrieval needs AI product machinery; this cluster's answer is an agent over plain files.
Distillate: From Notetaking to Neuralink (Contrary Research) — the 2023 pre-agent problem statement (information overload, manual-PKM limits, static→dynamic retrieval) beneath this pattern.
Capture-Organize-Distill-Express (CODE) — the CODE workflow this store runs; its Express stage (output as the pressure test) is the stage the retrieval-centric framings stop short of.
Organize by Actionability — PARA: the organize-stage method for keeping the store tethered to active work.
Distillate: Second Brain Explained for Engineers and Knowledge Workers — an independent (non-YouTube, non-Contrary) source converging on tool-agnosticism ("tool choice shifts ergonomics more than philosophy"), the storage-cheap/retrieval-expensive economics, and the AI-accelerates-but-judgment-stays-human division; adds the wiki-stabilization-gradient and complementarity claims above.

AI Second Brain

Claims

Related

Linked from