Record memory

Record Memory

Conversation context does not persist. When knowledge would save a future agent meaningful time, write a memory record immediately. Phase and slice wrap-up are mandatory prompts: scan notes.md, the active runtime phase sheet, audit findings, and fresh gotchas for durable guidance worth keeping.

Promote only durable guidance — repeatable workflows, sharp edges, invariants, subsystem facts. Not conversational state.

1. Choose the narrowest type that fits

fact (atomic checkable truth) · pattern (repeatable recipe / command sequence) · system (subsystem map + pointers, not a spec) · concept (stable mental model / terminology) · signpost ("start here" navigation) · thread (short-lived working set, expires fast).

⚠ A thread is hidden from find/retrieve until verified (SL-008 D6, §5). If you want a working loop to resurface by scope, prefer a durable type (pattern/system/concept) — or record the thread and verify it on a clean tree. An unverified thread shows only in list/show, never in scope ranking.

2. Record it

Record with doctrine memory record (ask --help for the flags; see using-doctrine.md for the verb model). It scaffolds a TOML + body under .doctrine/memory/items/; the born git anchor is captured automatically — do not hand-author it.

After recording

Suggested relations. record emits suggested relations on stderr when it detects high-confidence matches against existing memories. Review these and run doctrine link for matches you confirm — this builds the durable graph edges a future agent traverses.

Graph vs body links. Use [[relation]] edges for durable graph structure (typed, machine-traversable); use [[mem.…]] wikilinks in body prose for contextual "see also" pointers. Relations surface in retrieve --expand; wikilinks don't — they are for human readers.

--lifespan selection. Pick the narrowest lifespan that fits the knowledge, from most to least durable:

Lifespan	Rule of thumb
identity	Never ages — subsystem identity, invariant, canonical name
semantic	10:1 decay — design rationale, architecture constraints
procedural	3:1 decay — command recipes, build steps, workflows
episodic	Baseline — one-off findings, bug notes, session context
working	Fast decay — transient todo, short-lived hypothesis

Defaults to episodic when omitted. Narrow scope + narrow lifespan together keep retrieval relevant.

3. Scope it so it will be found

Scope a memory along four axes (the flags are in --help):

• path — exact file(s); strongest match.
• glob — subsystem relevance (e.g. src/auth/**).
• command — tied to a command flow (token-prefix).
• tag — stable categorization; do not overload tags as scope.

4. Set the risk axes

record defaults [trust] trust_level = "medium", [ranking] severity = "none", weight = 0. For risky or drift-prone memories, pass the flags at record time or update later:

doctrine memory record ... --trust high --severity critical

Or adjust after the fact:

doctrine memory edit <REF> --trust low --severity medium

• --trust = confidence: low (inferred, unvalidated) · medium (derived from reasonable context — default) · high (verified against code / specs / direct observation).
• --severity = how much it matters if wrong or ignored: critical · high · medium · low · none (default).

Calibrate honestly: creating a memory is authoring, not verification.

5. Attest if you verified it

doctrine memory verify <UID|KEY>

Stamps verified_sha against the working tree (refuses a dirty tree — no false attestation). Unattested memories read as lower trust.

Threads require this to surface at all. thread_expiry drops any thread that is not verified AND reviewed within 14 days from find/retrieve (SL-008 D6). record always writes unverified, so a fresh thread is invisible to scope ranking until you verify it — and verify refuses a dirty tree, so attest from a clean tree. Other types are never gated this way.

Holdback caution: retrieve suppresses low-trust ∧ high-severity memories (non-bypassable). A high-severity claim you have not verified will be held back until you raise its trust by attesting it — by design. Set trust to match reality.

6. Keep the body short and executable

Put "do X" steps in bullets with exact command snippets. Reference related artifacts and memories inline with [[uid]] / [[key]] (cheaper than relations). If the item would become an ADR or an evergreen spec, STOP — doctrine adr new or author a spec or ADR instead. Memory is a pointer/recipe layer, not canon.

7. Sanity-check surfaceability

Run doctrine memory find scoped to the context a future agent will actually query (see --help), and confirm the memory appears.

Exception: an unverified thread will not appear here even when recorded correctly (§1/§5) — verify it first, or check memory show instead.