video youtu.be Watch ↗

I Gave Claude Code a Permanent Memory

TL;DR

A 7-minute walkthrough (Taoufik) that rebuilds — a fourth independent time in this vault — the Karpathy-style second brain: two folders, raw/ (the inbox you write into) and wiki/ (the brain, where Claude Code links pages together), plus an index and log. Nothing here is new to the graph at the layout level; its value is as corroboration and in two mechanisms it makes concrete. First, ingestion is staged behind operator approval — a "watch" skill drops a report.md into raw/, dedups against the wiki ("already documented, doesn't duplicate"), and refuses to write to the brain until the operator says "ingest it now" (watch → report → approve → ingest → verify). Second, writes are wrapped in deterministic Python hooks: a session-start hook loads who-you-are, and a pre-write "brain-lint" hook checks the store is healthy enough to add to and flags orphan pages before anything lands. The opening frame crosses out "the bigger the better?" — a bigger graph is worse if it is mostly noise the agent can't use. The rest (typed templates per capture kind, /sweep to batch-process the inbox, a Dataview dashboard, Obsidian + Templater + Dataview as the stack, a "grill me" profile-setup skill) restates cluster claims.

Concepts introduced

AI Second Brain — the whole video is another independent instance of the store: raw/+wiki/ +index+log, credited to Karpathy, extraction via a "grill me" interview skill. Updated, not duplicated.
Knowledge-Graph Lint — the one genuinely additive node. The "brain-lint" pre-write hook: a deterministic structural-health check (orphan/relatedness) that gates writes to the store.
Knowledge Graph Retrieval — why the lint matters: the link structure is the index, so an orphan page is unreachable by the agent's graph-walk.
Agent Rituals — the session-start hook (load who-you-are) and pre-write hook are begin/before- write rituals delivered as deterministic code hooks rather than context-file instructions.
Validation-Gated Update — the brain-lint gate is its structural sibling: accept a write only if the graph stays connected, vs accept an edit only if a held-out score improves.
Multimodal Video Ingestion — the "watch" skill: an agent ingests a YouTube video, dedups it against the wiki, and stages a report for approval.
Evergreen vs Volatile Context — the "bigger graph ≠ better / too much noise" opening, at the store level: curate what gets in.

Held, not dropped (touched, no concept page yet):

Approval-staged ingest pipeline (watch → report → approve → ingest → verify) — a named, human-in-the-loop staging flow. Currently folded into AI Second Brain and Multimodal Video Ingestion as a best-practice claim; spin out as its own concept if a second source names the same five-stage shape.
/sweep as an on-demand batch consolidation command — process only the unprocessed raw entries into the wiki. Corroborates Incremental Indexing (do only what changed) and Offline Consolidation Cycle (batch inbox→brain); held rather than given its own page.
Typed capture templates (article/asset/book/brainstorm) via Obsidian Templater — the note-type analogue of Golden Templates' adapt-a-template move; held as a claim on AI Second Brain.
/save-chat (persist a conversation into raw/) and the Dataview dashboard / sources tracker — concrete surfaces; held.
The tool stack itself — Obsidian + Templater + Dataview + Notion kit; tool-specific, not concept-worthy.

Key claims

A second brain is two folders — raw/ (inbox) and wiki/ (agent-linked brain) — inspired by Karpathy. observation — the source's framing; a fourth independent convergence on the raw/wiki/index/log layout already held by AI Second Brain.
Bigger graph isn't better: noise and unusable context hurt the agent's retrieval. (principle, as the source asserts) — the crossed-out "the bigger, the better?"; restates Evergreen vs Volatile Context at the store level.
Don't write to the brain until the operator approves; dedup against the wiki at watch time. best practice — context: agent-driven ingestion where provenance/trust matter; the report sits in raw/ unprocessed until "ingest it now," and re-watching a documented source reports "already documented" instead of duplicating. → AI Second Brain, Multimodal Video Ingestion.
Wrap agent writes in deterministic code hooks — a session-start hook loads the profile; a pre-write hook validates and lints the store — so a stochastic write can't degrade the graph without passing a mechanical check. best practice — context: harnesses with pre-write hooks, unattended/frequent writes. → Knowledge-Graph Lint, Agent Rituals.
"Brain lint" checks the store is healthy enough to add to and flags orphan pages before writing. (observation → principle) — the source's description of the health gate; generalized as Knowledge-Graph Lint (orphans break reachability in a link-indexed store).
A template per capture type lets the agent fill a fixed structure instead of inventing one. best practice — context: an Obsidian+Templater store; classify type, instantiate that template. → Golden Templates (code-exemplar sibling), AI Second Brain.
/sweep batch-processes only the unprocessed raw inbox into the wiki. observation — incremental consolidation; → Incremental Indexing.
The ingested demo report asserts Anthropic launched "Routines" (schedule/webhook/API-triggered managed cloud sessions). (observation — the ingested source's claim, not adjudicated here) — shown on-screen inside the staged report.md; flagged for a later grounding pass, not verified.

Why this corroborates the second-brain cluster

This is the highest-frequency signal in the video, so it sets graph_relation. Taoufik's build is the fourth independent instance in this vault of the same store — raw/+wiki/+index+log, credited to Karpathy, with a "grill me" interview skill for extraction — after Nate Herk, Austin Marchese, and the Fable-5 LLM-wiki gist (see AI Second Brain). Convergence from four unrelated creators on the same layout, origin story, and extraction move is itself the finding: this is a stable pattern, not one author's idiosyncrasy. Because it corroborates, no concept is duplicated — the claims are appended to AI Second Brain with Taoufik as the new source.

Two threads are more than corroboration and are recorded as such. It builds on the cluster by making the approval-staged ingest pipeline (watch → report → approve → ingest → verify) and the deterministic write-hook delivery explicit — new best-practice detail on AI Second Brain and Agent Rituals. And it is novel in exactly one place: the brain-lint health gate, spun out as Knowledge-Graph Lint — a structural gate (keep the graph connected; no orphans) distinct from Validation-Gated Update's content-score gate and Input & Output Guardrails' model-I/O gate. That node lands squarely on firehose's own orphan-check discipline.

Illustrated walkthrough