Memory System
Cellium Agent implements a three-layer memory system for persistent context and knowledge storage.
Memory Layers
| Layer | Implementation | Description |
|---|---|---|
| Personality | personality.md | Static personality settings |
| Session | MemoryManager | Short-term context, bounded history |
| Long-term | FTS5 + Repository | Vector retrieval + hybrid recall |
API Vector Embedding
The long-term memory supports optional API-based vector embedding for semantic search:
- Model: OpenAI-compatible API (e.g. text-embedding-3-small)
- Dimensions: Auto-detected from API response (default 1536)
- Similarity: Cosine similarity
- Hybrid Recall: FTS5 full-text + vector similarity fusion
- Configuration: config/agent/embedding.yaml (disabled by default)
Schema & Category
Schema Types
| Schema | Purpose | Categories |
|---|---|---|
| general | General memory, session notes | general, user_info, command, project |
| profile | User profile | preference |
| project | Project-related | project |
| issue | Troubleshooting | troubleshooting, code |
Sensitive Information Control
Built-in detection and protection for sensitive data:
- Auto-Detection - Identifies private keys, AWS keys, GitHub tokens, API keys
- Redaction - Replaces sensitive values with [REDACTED]
- Write Blocking - High-risk sensitive info blocked from memory by default
- Classification - Sensitive memories are tagged for retrieval filtering
Security
Always review memory content before sharing. The system provides best-effort detection but may miss some sensitive patterns.