recall

/recall

Search Origin's memory by natural-language query. Returns matching memories ranked by hybrid vector + FTS search, then re-ordered by the agent if it helps.

Two phases

When a local model or API key is configured, the daemon can rerank and expand server-side. In local memory mode it cannot. The skill always does agent-side expansion and rerank itself — cheap, makes results good in both modes.

Phase 1 — expand the query (agent-side)

Before calling recall, rewrite the user's query into a more search-friendly form:

• Replace pronouns with the referent ("it" → the actual thing).
• Expand abbreviations the embedder is unlikely to know.
• Add the obvious synonym when the original term is too narrow (e.g.

"auth" → "auth OR authentication").

Don't over-expand. If the query is already specific, leave it alone. One recall call per /recall invocation — duplicate calls double embedding load and the merge step is rarely worth it. The daemon's own search_memory_expanded exists for the multi-query case; if it matters, use that endpoint instead of issuing parallel calls here.

Phase 2 — call the MCP tool

recall(query="<expanded query>", space=<inferred>, memory_type=<inferred>)

Inferences (do not ask the user):

• space: current working directory (e.g. ~/Repos/origin/... → "origin"),

the topic being discussed, or whatever space was mentioned in recent turns. Always pass when scope is known; if uncertain, run list_spaces later (post-PR-C) or omit.

• memory_type: only when the query itself names a type ("decision on X",

"lesson about Y", "preference for Z"). Otherwise omit and let hybrid search rank.

• limit: default 10. Use 3-5 for quick lookups, 10-20 for exploration.

Phase 3 — rerank (agent-side)

The daemon returns hits ranked by hybrid search. That ranking is good but not perfect — it doesn't know the user's exact intent.

Re-read the returned memories against the original query. Promote the ones that directly answer the question; demote ones that just share keywords.

Show the user the top 3-5 reranked hits. Surface the rest only if asked.

Phase 4 — render revision context (per result)

Each memory may carry revision fields: version, pending_revision, merged_from, last_delta_summary. Most memories are fresh (v1, none set) — render nothing extra for those. Only add a tag line when something meaningful is present.

Condition: emit the tag line when any of these holds:

• version > 1
• merged_from is non-empty
• pending_revision == true

Format — one compact line above the memory body:

<id>  v<N> (merged <K> memories)         ← merged_from has K entries
<id>  v<N>, pending revision against <id> ← pending_revision true
<id>  v<N> — <last_delta_summary>         ← version > 1, delta populated
<id>  v<N>                                ← version > 1, no delta

Rules:

• Merged takes precedence over pending_revision in the label.
• Omit — <delta> when last_delta_summary is empty or null.
• Skip the tag line entirely when version == 1 (or null) and no other

flag is set. Preserves current output for fresh memories.

When to use

• "What did I say about X?"
• "Do you remember the decision on Y?"
• Need a specific fact before continuing.

When NOT to use

• Broad session orientation → use /brief instead.
• Storing a new memory → use /capture.

Hint: write specific queries

"Alice database preference" finds more than "database stuff". The semantic matcher rewards specificity. If too many results return, add filters rather than making the query longer.