---
allowed-tools: Task
description: Save session learnings to cognitive memory
---

**IMPORTANT: Do NOT narrate your steps. Do all analysis silently. Your only visible output should be ONE of:**
- "Nothing new worth storing since the last save."
- "Saving N memories in the background." (followed by launching the agent)

## Process (do this silently)

1. **Find cutoff**: Scan for the most recent `memory_store` MCP call or `claude-memory store` Bash call. Only analyze conversation AFTER that point. If none found, analyze everything.

2. **Analyze**: Identify storable items after the cutoff — solutions, decisions, fixes, configs, patterns, insights. Include project names, technical details, rationale, before/after data.

3. **Gate**: If nothing after the cutoff is worth a memory (routine chat, minor reads, trivial refinements), say "Nothing new worth storing since the last save." and stop.

4. **Build summary**: Create a structured prompt for the agent:
   ```
   PROJECT: <name(s)>
   ITEMS:
   1. [type] Title / Tags / Importance / Content
   2. ...
   ```

5. **Launch agent**: Spawn in background with sonnet model:
   ```
   Task(subagent_type="memory-saver", model="sonnet", run_in_background=true,
        description="Store session memories", prompt="<summary>")
   ```

6. **Confirm**: Say "Saving N memories in the background."

## Guidelines

- Be thorough — capture everything worth remembering
- Don't duplicate memories already stored during the session
- Each item should be self-contained and useful on its own
- 1-6 items per session is typical; more is fine for large sessions
- Prefer specific, searchable titles over vague ones