Memory (MCP Memory Server)

MCP Server: mcp-memory

Automatic triggers via hooks

• Session Start: Relevant memories injected automatically based on project context
• Session End: Insights captured automatically when conversation ends
• Mid-Conversation: Pattern detection triggers memory retrieval (e.g., "what did we decide about...")

Manual Tools

When you need context the hooks didn't provide:

• retrieve_memory — semantic search by content
• recall_memory — natural language time queries ("last week", "yesterday")
• search_by_tag — find memories by project/topic tags

Storing Memories

Use store_memory when triggers below are met:

{
  "content": "One atomic fact or learning",
  "metadata": {
    "tags": "project-name,topic,category",
    "type": "note"
  }
}

Content Structure

✅ ONE fact per memory — never combine multiple facts ✅ Atomic memories — split by component/feature/concern ✅ Use descriptive tags for searchability ❌ No code blocks or documentation dumps

Good tags: foggyball-auth, scala-bloop, database-schema Bad tags: stuff, misc, temp

Automatic Storage Triggers

Store a memory when ANY of these occur:

1. User corrects me — their preferred approach/style
2. Non-obvious solution found — debugging insight that took effort
3. Project convention discovered — naming, structure, patterns not in docs
4. User expresses preference — explicit or implied ("I prefer X")
5. Custom tooling encountered — scripts, aliases, workflows unique to user
6. Architectural decision made — and its rationale
7. Recurring issue identified — something that keeps coming up
8. Domain knowledge learned — business logic, terminology, constraints

When NOT to Store

• Ephemeral session-specific details
• Information easily found in project files
• Trivial facts that won't aid future work
• Sensitive credentials or secrets
• Standard library usage (use Context7 instead)
• One-off tasks unlikely to recur

Quality Criteria (for HIGH scores ≥0.7)

The MCP Memory Service scores memories 0.0-1.0 using AI + usage signals.

Content Quality (DeBERTa classifier - 70% weight)

Example HIGH:

Python async: Use `asyncio.gather(return_exceptions=True)` for parallel API calls.
Prevents one timeout from cancelling siblings. Learned debugging webhook handler.

Example LOW:

async stuff - gather does something with exceptions?

Usage Signals (30% weight)

• Access frequency (40%) — Store things you'll search for
• Recency (30%) — Active knowledge beats stale
• Search ranking (30%) — Use specific, findable terms

Network Bonuses

Memories with ≥5 connections to others get +20% quality boost.

Quality Checklist

Before storing, verify:

• One atomic fact (not multiple combined)
• Includes context/rationale
• Contains searchable keywords
• Tagged with project + topic
• Would be useful if retrieved in 6 months