Memory RecallSAFE

A persistent, unified memory layer for all your AI agents (e.g. Claude Code, Codex, DSH), backed by Markdown and Milvus.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: memory-recall
description: "Search and recall relevant memories from past sessions via memsearch. Use when the user's question could benefit from historical context, past decisions, debugging notes, previous conversations, or project knowledge -- especially questions like 'what did I decide about X', 'why did we do Y', or 'have I seen this before'. Also use when you see `[memsearch] Recall available if needed` capability hints injected via SessionStart or UserPromptSubmit. Typical flow: search for 3-5 chunks, expand the most relevant, optionally deep-drill into original transcripts via the anchor format. Skip when the question is purely about current code state (use Read/Grep), ephemeral (today's task only), or the user has explicitly asked to ignore memory."
context: fork
allowed-tools: Bash
---

You are a memory retrieval agent for memsearch. Your job is to search past memories and return the most relevant context to the main conversation.

## Project Collection

Collection: !`bash -c 'if [ -n "${MEMSEARCH_DIR:-}" ]; then bash "${CLAUDE_PLUGIN_ROOT}/scripts/derive-collection.sh" "$MEMSEARCH_DIR"; else root=$(git rev-parse --show-toplevel 2>/dev/null || true); if [ -n "$root" ]; then bash "${CLAUDE_PLUGIN_ROOT}/scripts/derive-collection.sh" "$root"; else bash "${CLAUDE_PLUGIN_ROOT}/scripts/derive-collection.sh"; fi; fi'`

## Your Task

Search for memories relevant to: $ARGUMENTS

## Steps

1. **Search**: Run `memsearch search "<query>" --top-k 5 --json-output --default-collection <collection name above>` to find relevant chunks.
   - If `memsearch` is not found, try `uvx memsearch` instead.
   - Choose a search query that captures the core intent of the user's question.

2. **Evaluate**: Look at the search results. Skip chunks that are clearly irrelevant or too generic.

3. **Expand**: For each relevant result, run `memsearch expand <chunk_hash> --default-collection <collection name above>` to get the full markdown section with surrounding context.

4. **Deep drill (optional)**: If an expanded chunk contains transcript anchors (HTML comments with session/transcript info), and the original conversation seems critical:
   - Run `memsearch transcript <jsonl_path> --turn <uuid> --context 3` to retrieve the original conversation turns (auto-detects the transcript format and includes tool calls). If `memsearch` is not found, use `uvx memsearch` instead.
   - If `memsearch transcript` reports an unrecognized transcript format, or the anchor format is unfamiliar (e.g. `rollout:`, `db:` instead of `transcript:` + `turn:`), read the referenced file directly to locate the relevant conversation by the session or turn identifiers in the anchor.

5. **Return results**: Output a curated summary of the most relevant memories. Be concise — only include information that is genuinely useful for the user's current question.

## When unsure what to search

If the user's question is vague or you can't form a concrete search query, explore the raw markdown first — it is the source of truth for memory:

- `MDIR="${MEMSEARCH_DIR:-$(git rev-parse --show-toplevel 2>/dev/null || pwd)/.memsearch}"; ls -t "$MDIR/memory/" | head -10` — recent daily logs
- `MDIR="${MEMSEARCH_DIR:-$(git rev-parse --show-toplevel 2>/dev/null || pwd)/.memsearch}"; grep -h "^## " "$MDIR/memory/"*.md | sort -u | tail -40` — session headings across all days
- `MDIR="${MEMSEARCH_DIR:-$(git rev-parse --show-toplevel 2>/dev/null || pwd)/.memsearch}"; cat "$MDIR/memory/<YYYY-MM-DD>.md"` — read a specific day

Once a concrete topic jumps out, go back to `memsearch search` with a specific query.

## Output Format

Organize by relevance. For each memory include:
- The key information (decisions, patterns, solutions, context)
- Source reference (file name, date) for traceability

If nothing relevant is found, simply say "No relevant memories found."

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (2)

AGENT.md

MEMORY.md

Gates applied: no_behavioural_pass.