Memory Recall

Drives the scoped recall ladder over the PGLite/pgvector memory store: search for a candidate, expand it for surrounding context only if needed, drill into the raw transcript only if exact wording is needed. Stops at the lowest rung that answers.

Outcome

• A sourced answer following the Found / Partial / Absent pattern from docs/memory-retrieval.md
• The answer cites the lowest rung that answered (search hit, expanded chunk, or transcript window) — never a raw file read
• No scope widening mid-ladder: every rung in one recall uses the same scope flags

Context Needs

None beyond what each rung's CLI already returns. Don't read context/memory/ or context/transcripts/ files directly — always go through the CLI.

Step 1: Tier 0 Recap (one line)

Tier 0 (context/MEMORY.md + today's daily log) is already loaded at session start per AGENTS.md. If the answer is already there, answer directly — do not invoke this skill's ladder. Only proceed to Step 2 if Tier 0 has nothing or is too thin to answer confidently.

Step 2: Search

From command-centre/:

npm run memory:recall -- "<query>" --system --json

Swap --system for --team <id> / --client <slug> / --user <id> depending on whose memory you're answering for (each also includes the system baseline). Read the top hits. Note the chunk_id of the best candidate.

If search returns nothing in scope, stop here — go to Step 5 with an Absent/Partial answer. Do not escalate rungs on an empty result.

Step 3: Expand (only if the best hit is too short to answer confidently)

npm run memory:recall -- --expand <chunk_id> --system --json

Use the exact same scope flags as the Step 2 search call. This returns the surrounding source context (neighbouring chunks from the same source). If this answers the question, stop here.

Step 4: Transcript (only if exact wording, a command, or an output is needed and expand wasn't enough)

npm run memory:recall -- --transcript <chunk_id> --system --json

Again, the exact same scope flags as Steps 2-3. This returns a window of raw conversation turns around the one the chunk summarizes.

Step 5: Answer

Stop at the lowest rung that answered. Apply the Found / Partial / Absent citation pattern from docs/memory-retrieval.md:

• Found — answer + inline source citation + temporal context (flag if >14 days old)
• Partial — what's known + what's missing + where you looked + temporal gap
• Absent — state what was checked (MEMORY.md, daily logs, semantic search across indexed sources) and that nothing matched

Rules

Updated automatically when the user flags issues. Read before every run.

• Stop-early discipline. Don't drill to expand or transcript "just in case" — only escalate when the current rung's content is genuinely too short or imprecise to answer.
• Identical scope flags across rungs. Once a recall starts with --system (or --team/--client/--user), every subsequent expand/transcript call in that same recall uses the identical flags. Never widen scope mid-ladder.
• Never bypass the CLI. All three rungs go through npm run memory:recall -- (which wraps memory-search.cjs / memory-expand.cjs / memory-transcript.cjs). Never query the database or read context/transcripts/ files directly.
• Null/empty means stop and say so. If a rung returns null or no matches, that is a real answer (not found in scope) — report it as Partial/Absent rather than guessing or fabricating a source.
• Explain scope in plain terms. You only ever see your own memory plus your team's plus the system baseline — never another client's or another user's private memory. If a result seems surprising in scope, say so plainly rather than implying broader access.

Self-Update

If the user flags an issue with how recall escalates rungs or cites sources, update the ## Rules section in this SKILL.md immediately with the correction and today's date.

Graceful Degradation

• Memory backend unavailable (exit code 3 from any rung) → tell the user to run bash scripts/setup-memory.sh, don't retry silently
• Empty result at any rung → Absent or Partial response per docs/memory-retrieval.md, never fabricate a source
• Ambiguous scope (unclear which tenant to search as) → ask the user rather than guessing a scope flag