How mnemoclaw/immune works
Immune is an adaptive memory system for AI-agent prompts: it retrieves positive strategies before generation, scans generated output for known or new errors, and updates memory from results. The README documents the overall workflow and capabilities; sampled adapter excerpts confirm hybrid retrieval, deduplication, session logging, and persistent storage. The adapter’s full command wiring and the scan instructions were not sampled, so uncertain implementation-level links are omitted.
Main entry points: immune-adapter.js JSON memories · immune-inject.js Strategy injection · immune-scan.md Immune scan
Components 19
Prompt preparation
- Prompt hook
immune-inject.js - Domain detection
immune-inject.js - Strategy injection
immune-inject.js
Scan and learning
- Immune scan
agents/immune-scan.md - Error correction
agents/immune-scan.md - Pattern learning
immune-adapter.js - Pattern dedup
immune-adapter.js - Output scoring
immune-adapter.js - Session logging
immune-adapter.js
Memory retrieval
- Hybrid retrieval
immune-adapter.js - Local embeddings
immune-adapter.js - FTS4 search
immune-adapter.js - RRF ranking
immune-adapter.js
Memory state
- Memory housekeeping
immune-adapter.js - JSON memories
immune-adapter.js - SQLite index
immune-adapter.js
Other components
- Agent user
- Output generation
- Compatible model
Connections 23
- Agent user → Prompt hook: submits prompt
- Prompt hook → Domain detection: detects domains
- Domain detection → Strategy injection: selects strategies
- Strategy injection → JSON memories: reads strategies
- Strategy injection → Agent user: injects context
- Agent user → Output generation: requests output
- Output generation → Immune scan: scans output
- Immune scan → Compatible model: checks errors
- Immune scan → Error correction: reports errors
- Error correction → Pattern learning: learns patterns
- Pattern learning → Pattern dedup: checks duplicates
- Pattern dedup → JSON memories: updates memories
- Hybrid retrieval → Local embeddings: ranks semantically
- Hybrid retrieval → FTS4 search: finds keywords
- Local embeddings → RRF ranking: returns ranked matches
- FTS4 search → RRF ranking: returns ranked matches
- RRF ranking → JSON memories: loads patterns
- FTS4 search → SQLite index: queries index
- Pattern learning → SQLite index: syncs index
- Output scoring → JSON memories: updates scores
- Session logging → JSON memories: writes context log
- Session logging → SQLite index: indexes session
- Memory housekeeping → JSON memories: archives patterns