| name | project-memory |
| description | Use when about to save or recall anything durable: before writing MEMORY.md, auto-memory, or CLAUDE.md, before answering from a stale summary, or before web-searching what a prior session already answered. Covers store, recall, curation. Do NOT use for raw transcript replay. Never stores secrets. |
Project memory
DURABLE FACTS LIVE IN fact_store, NOT IN MEMORY.md OR CLAUDE.md.
RECALL BEFORE RE-DERIVING; STORE WITHOUT BEING ASKED.
Announce: "Using tracedecay:project-memory to <recall/store/curate>."
Route the moment
| Moment | Action |
|---|
| Need a prior decision/preference/pitfall | tracedecay_fact_store action:"search" (query, min_trust) — before web search, before asking the user |
| Prior conversations, not facts | tracedecay_message_search (query, limit) — this skill owns FTS→fact; raw replay/scoped grep → tracedecay:managing-session-context |
| A durable decision/correction/pitfall just surfaced | tracedecay_fact_store action:"add" (content, category, tags, trust) — proactively, do NOT wait to be asked, and do NOT write MEMORY.md instead |
| User rates a recalled fact | tracedecay_fact_feedback (helpful/unhelpful) |
| A recalled fact you were shown helped or misled you | tracedecay_fact_feedback on its fact_id (helpful/unhelpful) — don't wait to be asked |
| User asks to clean/merge/delete memory | Curation flow below |
Trust calibration for adds: 0.85+ independently verified decisions, ~0.7
ordinary well-sourced facts, ~0.5 plausible-but-uncertain. The add path
already rejects secrets and reports near-duplicates/conflicts — act on those
flags; never rephrase a rejected secret to bypass filtering.
Rate what you recall: any fact_id shown in tracedecay_context's Memory
Matches (or returned by fact_store search) that materially helped or misled
you should get tracedecay_fact_feedback (helpful/unhelpful) the moment you
act on it — proactively, without waiting for the user, since recalled facts are
almost never rated and feedback is how trust is earned.
Do NOT capture: secrets/credentials/PII, transient errors,
environment-specific failures, one-off narratives, task progress, or
soon-stale session outcomes — those belong to session transcripts.
Curation (mutation — parent agent only, approval required)
Read references/curation.md for the full protocol:
read-mostly inventory → native dry-run (tracedecay memory curate) →
candidate buckets → narrow apply (fact_store add/update/remove) → read-only
verify (tracedecay_memory_status for counts/health). Hard rules that always
apply:
- Deletion is permanent (no soft-delete, no undo). Explicit approval
immediately before every
remove, hard delete, or merge-loser removal,
showing fact id, content summary, and reason. Prefer update/merge when
provenance should survive.
- Subagents may inspect and recommend only — never let a subagent call
add/update/remove/feedback, apply curation ops, or run memory repair.
- Do not lower trust merely for age; cite newer evidence or a contradiction.
If tools are deferred or MCP fails
- Deferred: one ToolSearch call —
select:tracedecay_fact_store,tracedecay_message_search,tracedecay_fact_feedback.
- MCP error:
tracedecay tool fact_store --action search --query … (see
tracedecay:using-the-cli). An MCP failure is not a reason to write
MEMORY.md — the CLI reaches the same store.
Deliverable
Recall: the prior context/decisions found, with source and trust — or an
explicit "no stored fact matches", after which storing the fresh answer is the
default next step. Store: the fact id(s) written and any duplicate/conflict
flags handled. Curate: facts changed/skipped, approvals obtained,
verification result. Report any tracedecay_metrics: line.
