StateRoom 01 of 13
It remembers, and keeps its memory clean.
A self-maintaining graph of the people, projects, and dates in your world. New facts supersede old ones, everything stays searchable, and every entry is yours to inspect.
Entity graph: you know Ana, who works on Project Atlas, due Mar 15. You live in Vienna.
mem_5e07c2 You live in Vienna
Recreation of the Dashboard memory graph, with sample data.
AWatch one memory, written to forgotten
The life of one memory.
Written, linked, scoped, found, corrected, and deleted. Each step is a row you can see.
content
Met [[Sarah Chen]] at the [[Atlas]] design review.
category=fact · scope=user
Linked on write
- Sarah Chenconcept · mentions
- Atlasconcept · mentions
Each target is upserted as an entity (default kind concept) with a mentions edge from the entry.
A memory.write call with wikilinks, and the entities it links. A switch shows the alias form.
Step 1 of 5: Write it down
01 / 05 · Write it down
Each entry is linked to the people, projects, places, and dates it mentions. Wrap a name in [[double brackets]] and it is linked as soon as the entry is written; a fast-tier model also scans every new entry for names in the background. The alias form [[Sarah Chen|Sarah]] links the canonical entity and keeps “Sarah” as display text only.
BMechanism
Outside the window, in a store you can read.
Long-term entries carry a category (fact, preference, or observation) and a scope. Daily logs carry what happened. Both live in your database, not in the model.
- A self-maintaining entity graph
- Sophon links the people, projects, places, and dates it records into a graph it maintains itself, so “Ana works on Project Atlas, due March 15” is one connected record, not three loose notes. The agent can traverse it instead of rereading every entry.
- Facts that stay current
- When something changes, the new fact supersedes the old one. The history is kept, but the assistant acts on what's true now. No stale context quietly steering the answer.
- Two scopes, cleanly separated
- Shared memory holds what's true about you and your world; agent memory holds how a particular agent should behave. You can see which is which, and edit either.
- Find anything, two ways
- Hybrid search combines semantic meaning with exact keywords. It works on every tier. Personal ships with a built-in vector store, and Sophon prefers an embedder on your own machine. A cloud embedder needs an admin's consent first. Until an embedder is ready, search uses keywords.
- Daily logs, then consolidation
- Each interaction adds a timestamped line to the daily log, and the agent reads the last two days on every turn. MemoryConsolidationJob, a cron job that runs daily at 03:00 UTC by default, promotes durable facts from recent logs into long-term entries through the same write pipeline as memory.write.
- Six tools for the agent
- memory.write, memory.search, memory.list, memory.forget, memory.graph (overview, neighbors, path, or cluster), and memory.diagram, which draws the graph as a Mermaid flowchart in chat.
CSpecifics
What the model receives, and who can see it.
Per turn
- Shared memory (user scope)
- 50 entries max
- Agent memory (this agent)
- 50 entries max
- Recent activity
- 2 days of daily logs
- Chat history
- 100 messages
The context assembler fetches shared entries and the active agent's own. When memory outgrows the caps, entries are ranked against the current message and the most relevant float up.
Isolation
- Per user
- Every query filters by the authenticated user ID. Semantic search filters by user and agent on the server before returning results.
- Per agent
- Agent-scoped entries are visible only to the owning agent.
- Per tenant
- On Enterprise, EF Core global query filters enforce tenant boundaries on every query.
- No secrets
- The credential vault and memory are separate stores. API keys and OAuth tokens never land in memory and never reach an LLM prompt.
The graph explorer
- The whole graph at once
- Memory, then Graph, opens every entity, entry, and link in a force-directed layout, up to a node cap: Compact 100, Balanced 250 (the default), Dense 500, or Maximum 1000. A banner says when the cap clips your graph.
- Filters that dim, not hide
- Search dims unmatched nodes and keeps their neighbors vivid. Kind chips, a date window of 7, 30, or 90 days, and a co-mention overlay refine the view.
- Duplicates, reviewed by you
- Candidate pairs come from exact name or alias collisions and from trigram similarity at 0.85 or above. Merge moves links and aliases onto the winner; Not a duplicate remembers the pair.
- Export as Mermaid
- From the export dialog, from memory.diagram in chat, or from GET /api/memory/graph/mermaid, with labels sanitized against the memory text they come from.
Limits
What memory doesn't do.
None of this is learning. No training, no fine-tuning, no reinforcement learning, no weight updates. Consolidation is a cron job writing rows, and you can read every one.
The thesisThe model sees a budget, not everything: up to 50 user entries, 50 agent entries, 2 days of daily logs, and 100 messages. Past the caps, entries are ranked against the current message.
Context assemblyUntil an embedder is ready, search is keyword-only. A cloud embedder that has not produced vectors yet embeds nothing until an admin consents.
EmbeddingsA local-only chat on a host with a cloud embedder recalls by keyword only, and its memories are still embedded later by that host embedder, so their text does reach the provider.
EmbeddingsThe embedded vector store is single-node. Multi-replica deployments need Qdrant, which is recommended above about 100,000 vectors.
Storage and semantic searchEntity-to-entity relationships are not stored rows. The co-mention overlay infers them from entries that mention both, on screen only.
Co-mention overlayDaily logs are not kept forever: by default, logs older than 30 days are removed after each consolidation sweep.
Consolidation and retention