Memory
Your delegate remembers you — not just what you said in this conversation, but things you’ve shared over time. The more you use it, the better it understands you.
How it works
Memory runs quietly in the background. As your conversation unfolds, the delegate is paying attention — picking up facts, preferences, and context without you having to spell them out. By the next conversation, it already knows them.
You don’t need to say “remember that I…” or “last time I told you…”. Relevant things from past conversations surface automatically when they’re useful.
Memory gets stronger over time
Something you mention once is noted lightly. Something that comes up again and again becomes well-established. This means a random throwaway comment won’t override something your delegate has learned to know well about you.
When similar memories start to pile up, the system consolidates them — deciding whether to reinforce, merge, or discard — so your memory store stays clean and meaningful rather than filling with noise.
Memory and background awareness
Consolidation, conflict, and override outcomes enter the Inner Context Plane as typed wake signals. Brand-new inserts remain observable without waking the agent. Passive retrieval also enriches every base-agent activation, including activations started by timers and other inner signals. The activation itself is the retrieval cue, so learned instructions can surface without a timer-specific workflow. The memory subsystem reports what happened; the base agent applies its prompt and recalled memories to decide whether any action is appropriate.
When consolidation wakes the agent, the conversation timeline shows the full signal batch, recalled memories, subsequent tool activity, and final outcome as expandable internal entries around the assistant response. Settings > Memory Insights remains a memory-store diagnostic view, not the primary Inner Context Plane history.
Idle time creates an incremental memory checkpoint rather than ending the conversation. Each checkpoint extracts only user and assistant turns added since the previous checkpoint, so memory can learn regularly while the user returns to the same continuous timeline hours later.
Native adaptive memory
- Adaptive memory is always on. It builds up a picture of you on your own device, stored privately.
- Retrieval, reinforcement, deduplication, and consolidation all run through the native memory system in
src/memory/.
Viewing and managing your memories
You can see what your delegate remembers about you at any time:
- Ask directly — say something like “what do you remember about me?” and it will tell you.
- Use the UI — go to Settings to browse and delete individual memories.
- Ask it to forget — say “forget that I…” and it will remove that memory.
Technical details
- All memory code lives under
src/memory/. - Adaptive memory extracts facts in-process and stores them on disk (
runtime-data/). - Conversation bus — a real-time pub/sub that lets the memory subsystem listen to assistant turns and extract memorable facts asynchronously after each turn.
- Deduplicator (
src/memory/deduplicator.ts) — suppresses near-duplicate inserts. Unit-tested vianpm run test:unit. - Local memory config:
runtime-data/memory-config.json, editable viaGET/PUT /memory-config.
| Endpoint | Purpose |
|---|---|
GET /api/memories |
List stored memories |
DELETE /api/memories/:id |
Remove a specific memory |
GET /api/memories/insights |
Runtime diagnostics (dedup, store stats, recent memory runtime events) |