NAOMS · building block
Agent memory
A personal knowledge graph that agents remember with, encrypted by default.
Agent memory is a graph database: every remembered fact starts as a signed chain event and becomes a graph node, one of fourteen kinds such as episode, fact, preference, or belief. Related memories are consolidated and scored for reliability over time, and agents pull from it with search and recall calls instead of a flat log. It gives an agent continuity across sessions grounded in your own history.
What a person notices
- A coding or personal agent remembers what it learned in an earlier conversation, so the person doesn’t repeat themselves.
- What the agent remembers stays on the device by default, encrypted like everything else.
- If the agent gets something wrong, there’s a record of what it claimed and when, not a silent, unexplained change of mind.
What you get as a developer
Agent memory is the knowledge graph specialised for what an agent needs: something it can store into, search, consolidate, and be honest about when it turns out to be wrong.
- Storage, search, consolidation and contradiction detection are separate, real stages, not one flat log: consolidation merges related memories over time, and contradiction detection flags when a new memory conflicts with an older one instead of silently overwriting it.
- An agent can promote a memory claim onto its own identity chain, then later re-attest, downgrade, or honestly forget that promotion with a stated reason, so a change in what an agent claims to know is recorded rather than dropped.
- A source-reliability score and retention policy travel with each memory, so an agent can weigh how much to trust something it remembers, not just whether it remembers it at all.
- Whether content from a person’s inbox is allowed to enter the agent’s knowledge surface is a separate, explicit setting, off by default.
store a memory: memory.store { text, source, tags }
search memories: memory.search { query }
promote to identity: memories.promote { memoryId }