| name | recall |
| description | Semantic search across verbatim Claude Code chat history from every session and project. Backed by the local claude-echoes server (pgvector + nomic-embed). Use when the user asks "when did we/I...", "find that conversation about...", "what did I say about...", "recall...", or "/recall". |
| user-invocable | true |
/recall — Semantic Recall Over Verbatim Chat History
Queries the local claude-echoes server for past Claude Code messages matching a natural-language query. Returns verbatim content with project, role, date, and session context.
Server: http://localhost:8088 by default. Override with ECHOES_URL env var for remote deployments.
Quick reference
| Command | What it does |
|---|
/recall <query> | Top 10 semantic hits across all projects |
/recall <query> --project <name> | Filter to one project |
/recall <query> --days <N> | Only messages from last N days |
/recall <query> --role user | Only user prompts |
/recall <query> --role assistant | Only assistant responses |
/recall <query> --limit <N> | Change result count (default 10) |
Execution
Step 1 — Parse the query. Extract the natural-language query and any --project, --days, --role, --limit flags.
Step 2 — Call the search endpoint.
curl -sG "${ECHOES_URL:-http://localhost:8088}/search" \
--data-urlencode "q=<the query>" \
--data-urlencode "limit=<limit or 10>" \
[--data-urlencode "project=<project>"] \
[--data-urlencode "role=<role>"] \
[--data-urlencode "days=<days>"]
If the server is unreachable, say so plainly. Do not fall back to guessing or summarising from your own context.
Step 3 — Present results. For each hit, show:
- Similarity score (3 decimals)
- Project name + role + date (convert UTC to user's local timezone)
- First 200 chars of content as a snippet, with ellipsis if truncated
- Session ID in
[abbreviated] form (first 8 chars) so the user can ask for full context
Format example:
0.731 azureprep user 2026-04-08 15:09 [ada9a195]
"I would like semantic search across verbatim chat history..."
0.684 cosmos-collective assistant 2026-04-05 22:14 [b831cc02]
"The partition key for the users container must be /email..."
Step 4 — Offer follow-up. After showing results, ask if the user wants the full surrounding conversation for any hit. If yes, call:
curl -s "${ECHOES_URL:-http://localhost:8088}/session/<session_id>"
And render the full message list in chronological order.
When NOT to use recall
- The user is asking about the current session — that's already in your context
- The answer is in code or git history — read files or
git log instead
- The user wants a project summary — recall is for specific conversations, not overviews
- The query mentions "today" or "right now" without specifying historical intent
Triggers
Use this skill when the user says any of:
/recall ...- "recall when I/we..."
- "find that conversation about..."
- "what did I say about..."
- "when did we fix/build/discuss..."
- "do you remember when..."
- "pull up the session where..."
