recall

Recall — Episodic Search Over Past Sessions

Law 1 says research before executing. The cheapest research is your own history: the exact error you are staring at may have been solved three sessions ago. observations.jsonl already records every tool call, but an append-only log is not searchable. Recall turns that log into a ranked, queryable memory.

When to Activate

• Before tackling a problem that feels familiar ("haven't I seen this error before?").
• Before a risky or destructive operation — check whether a past attempt failed.
• Before reading large files from scratch — a past session may already summarize the relevant facts.
• When onboarding into an unfamiliar area of the codebase that you have touched before.

Core Concept

Recall builds an in-memory BM25 index over the observation rows and answers a query with the most relevant past activity, newest-first on ties:

ci_recall query="permission denied push"
ci_recall query="jq command not found" k=3
ci_recall query="auth login" since=7d

Each result is a past tool call with a redacted snippet, a relevance score, and a timestamp.

Privacy

Snippets are passed through a secret redactor before they are surfaced. AWS access keys, JWT-shaped triplets, bearer tokens, KEY/SECRET/TOKEN/PASSWORD assignments, and long hex strings are masked. The observation log already caps output at 200 characters; redaction is the second layer.

Limitations

• Lexical, not semantic. A query for "login" will not surface activity that only ever said "authentication". Search with the vocabulary that actually appeared in the tool calls, or try several phrasings.
• Scoped to the captured history. Recall only knows what the hooks recorded. Thin-schema rows (no input/output, emitted when the Node observer is not wired) contribute little signal.
• In-memory rebuild per query. Dependency-free and fast at current volumes; a node:sqlite FTS5 index is a planned follow-up if the log grows past ~100k rows.