How It Works
- On each message, relevant memories are retrieved by cosine similarity and injected as context
- The harvester periodically extracts facts from conversations and stores them as embeddings
- The cleaner periodically deduplicates, merges, and normalizes stored memories
Scopes
Memories are scoped to control visibility:| Scope | Description |
|---|---|
user | Personal facts about a specific user |
agent | Context specific to an agent |
global | Shared knowledge across all agents |
Pinned Memories
Memories can be pinned — pinned memories are always injected into the system prompt regardless of similarity score. Use pinning for critical context that should always be available.Configuration
See the memory configuration section for all options including embedding model, harvester schedule, and similarity thresholds.Tools
Agents have built-in tools for memory management:| Tool | Description |
|---|---|
save_memory | Store a new memory |
search_memories | Semantic search across memories |
forget_memory | Delete a specific memory |
list_memories | List memories with optional filters |