Save Memory Skill

Persist high-signal, non-bloated project memory into the project-memory MCP server.

This skill strictly partitions memories into 3 categories (ledger, sidecar, journal), enforces crisp summaries (no raw data or stack traces), and mandates documenting learnings from mistakes and errors encountered or avoided.

The 3 Memory Categories

[!IMPORTANT] Journal Entries Require BOTH Session Summary AND Learnings: A journal entry captures the summary of the current session with learnings. Never save a journal entry with just learnings alone! An agent reading the journal in a future session needs the context of what was accomplished/transpired in that session (summary) alongside what lessons and mistakes were learned/avoided (learnings).

Anti-Bloat Rules (STRICT)

1. NO Raw Code Dumps: Never paste entire source files or diffs into summary. Use 2–4 concise bullet points.
2. NO Stack Traces or Logs: Never dump raw terminal outputs, stack traces, or base64 blobs into memory.
3. Session Summary WITH Learnings for Journal: For journal, summary captures the session's work and progress, and learnings captures the lessons learned and mistakes avoided. Both are required and must not be duplicates.
4. Mandatory Learnings: Every save MUST detail the specific mistake, trap, or wrong assumption that was made or avoided, why it happened, and how to prevent it. Never write placeholder text like "none" or "n/a".
5. Summary Limit: Keep summary under 1,500 characters.
6. Title Limit: Keep title under 150 characters, descriptive and searchable.

Workflow

Step 1 — Classify the Memory

Select the category:

• Is this a permanent project truth or architecture decision? → ledger
• Is this transient context about what is currently being worked on? → sidecar
• Is this a session summary with lessons and takeaways? → journal

Step 2 — Construct Summary and Extract Learnings

For a journal entry, construct both parts:

1. summary (Current Session Summary): What was worked on, accomplished, or altered during the current session?
2. learnings (Takeaways & Mistakes Avoided): What key lessons were learned, what errors or traps were encountered, and what rule ensures future sessions succeed?

Step 3 — Invoke save_memory via MCP

{
  "name": "save_memory",
  "arguments": {
    "category": "journal",
    "title": "Session: Scaffolded SQLite MCP Memory Server & Hybrid Search",
    "summary": "Completed initial scaffolding of project-memory MCP server with node:sqlite DatabaseSync and verified 40% semantic hybrid search test suite.",
    "learnings": "Native C++ node addons fail on Windows sandbox without build tools; standardizing on built-in node:sqlite DatabaseSync avoids node-gyp build failures completely.",
    "tags": ["sqlite", "mcp", "session-recap", "hybrid-search"]
  }
}

Step 4 — Verify Output

Ensure the returned response contains:

• success: true
• Clean MemorySummary object with ID and timestamps
• Confirm that NO vector embeddings or raw payloads are returned