majiayu000 In Memoria: Persistent Codebase Intelligence
In Memoria is an MCP server that learns your codebase patterns once, then exposes that intelligence to AI agents persistently. Instead of re-analyzing code on every interaction, it maintains a semantic understanding of your architecture, conventions, and decisions.
Core Concept
Setup → Learn → Verify → Serve. After that, AI agents query persistent intelligence without repeated parsing.
Quick Start (5 minutes)
# 1. Configure for your project
npx in-memoria setup --interactive
# 2. Build intelligence database
npx in-memoria learn ./src
# 3. Verify it worked
npx in-memoria check ./src --verbose
# 4. Keep it fresh (optional but recommended)
npx in-memoria watch ./src
# 5. Expose to agents via MCP
npx in-memoria server
When to Use
✅ Use In Memoria:
- Building long-lived AI agent partnerships (Claude, Copilot, etc.)
- Projects where consistency across sessions matters
- Teams wanting shared codebase intelligence
❌ Skip it:
- One-off analysis (use
npx in-memoria analyze [path]directly) - Simple projects agents can read directly
The 5 Core Commands
| Command | Purpose | When |
|---|---|---|
setup --interactive |
Configure exclusions, paths, preferences | First time only |
learn [path] |
Build/rebuild intelligence database | After setup, major refactors |
check [path] |
Validate intelligence layer | After learn, before server |
watch [path] |
Auto-update intelligence on code changes | During development (optional) |
server |
Start MCP server for agent queries | After check passes |
Key difference: learn builds persistent knowledge. analyze is one-time reporting only.
What Agents See
When connected, agents can query:
- Project structure - Tech stack, entry points, architecture
- Code patterns - Your naming conventions, error handling, patterns used
- Smart routing - "Add password reset" → suggests
src/auth/password-reset.ts - Semantic search - Find code by meaning, not keywords
- Work context - Track decisions, tasks, approach consistency
Troubleshooting
| Issue | Fix |
|---|---|
| Learn fails | Verify path is correct; check file permissions |
| Check reports missing intelligence | Run learn [path] again |
| Agent doesn't see new code | Is watch running? Start it: npx in-memoria watch ./src |
| Server won't start | Run check --verbose first; if issues, rebuild: rm .in-memoria/*.db && npx in-memoria learn ./src |
| Multiple projects conflict | Use server --port 3001 (or different port per project) |
Performance Notes
- Small projects (<1K files): 5-15s to learn
- Medium (1K-10K files): 30-60s
- Large (10K+ files): 2-5min
If learning stalls (>10min), verify you're not indexing node_modules/, dist/, or build artifacts—use setup's exclusion patterns.
Key Principles
- Local-first - Everything stays on your machine; no telemetry
- Persistent - One learning pass; intelligence updates incrementally with
watch - Agent-native - Designed for MCP; works with Claude, Copilot, and any MCP-compatible tool
- Pattern-based - Learns from your actual code, not rules you define
Deployment Pattern (3 terminals)
# Terminal 1: One-time setup
npx in-memoria setup --interactive
npx in-memoria learn ./src
npx in-memoria check ./src --verbose
# Terminal 2: Keep intelligence fresh
npx in-memoria watch ./src
# Terminal 3: Expose to agents
npx in-memoria server
# Now agents (Claude, Copilot, etc.) have persistent codebase context
See GitHub for full API docs and agent integration examples.