How it works

In the chain (links ② + ③). This page is the engine of the whole product: *because the code is the truth, the AI stamps the meaning at write-time (②), which is exactly why reading it back is a 0-token lookup (③).* Free reading is what makes the 6 dimensions and search-inversion affordable on every edit. See Why it all connects.

Write-time stamping, not read-time guessing (ADR-001)

The foundation. Other tools (and raw RAG) try to understand the code at read time — every query re-derives meaning with an expensive model. recall flips it: the expensive AI writes the meaning down once, at commit time, while the context is already in its head. Reading it back is then a dumb, fast database lookup.

  write-time (once, ~free):   AI stamps the WHY + anchors onto a knowledge edge
  read-time (millions, 0 tok): SQLite + FTS5 returns the finished edge, sub-ms

The .mind index

recall init builds .mind/index.db (git-ignored), holding:

• the code map — symbols, files, importance, dependency edges (tree-sitter,

no model);

• commit facts — immutable, can't go stale;
• knowledge — lessons, decisions/ADRs, tasks: the claim-bearing notes that CAN

drift, and so are the only things drift-tracking watches.

The three (now four) tracks

A query returns independent tracks, each ranked on its own axis so a loud one never buries another (ADR-028): code (where, by importance), knowledge (why), blast radius (what breaks), open tasks (standing intent). Search-inversion (resolve) adds a vocabulary track on top.

Anchors and edges

A stamp carries anchors — the rare, load-bearing terms (a migration number, a symbol name, an ADR id) by which it should be found — and optionally typed edges (supersedes, guarded_by, depends_on, …). The intelligence lives in the edge; the retriever stays dumb.

Why the read path is 0 tokens

There is no LLM in the read path — it's FTS5 ranking over text the AI already wrote. That's the whole economic argument: orienting before an edit is free, so skipping it never saves tokens; it only loses the why, the open tasks, and the blast radius.

Drift can't hide

Because there's one truth (the code) and the notes are SHA-anchored, a note whose file moved on is flagged 🟡/🟠 — never silently trusted, never silently rewritten. See Governance & drift.

Measured, not claimed

The numbers (retrieval quality, token savings vs grep-and-read) are reproducible via the scripts in experiments/ and recorded in docs/benchmarks.md. Search-inversion in particular: blind grep of a hallucinated term finds the real symbol 0/12 of the time — that gap is what resolve closes.