Cognition
Cognition is where long-term Memory lives in Nimi. When an Agent remembers your name, your preferences, or something you told it last month, that memory came from Cognition.
Cognition is its own owner inside the platform. Runtime runs the conversation; Cognition keeps what is worth remembering. The two are connected by a narrow, typed Bridge, and neither side absorbs the other.
What This Means For You
- Your Agent remembers you across sessions. Close the app, come back next week, and the same Agent still knows what you told it, as long as it was worth keeping. Memory belongs to the Agent itself, not to one conversation or one world, so it stays with that Agent.
- Memory is private to each Agent. Two Agents on the same account keep separate memories, and neither can read the other's. There is no shared account profile behind them.
- You can correct an Agent. When you say "I don't live in Berlin anymore," the corrected fact becomes current right away, and the old one leaves default recall, kept only as bounded history.
- You can make an Agent forget. An explicit forget is a durable barrier: restarts, retries, index rebuilds, and replays cannot bring the forgotten memory back.
- Secrets don't become memories. Passwords, tokens, credentials, and content marked "do not remember" are rejected before they can be retained.
- Memory knows how it knows. Each memory carries honest provenance: whether you said it explicitly, whether it was inferred, or whether it was consolidated from earlier memories.
Memory is also allowed to be absent. If recall is still building, unavailable, or simply finds nothing, the conversation continues normally with zero memories. A read hiccup never fabricates results and never blocks a turn. Writes are stricter: a correction or forget is only reported done after Cognition has durably committed it.
What Cognition Owns
- Each Agent's long-term Memory: one opaque, Agent-private logical bank bound to one Agent. Not an account profile, not a workspace bank, and not something an App can query directly.
- Memory truth and lifecycle: admission, provenance, correction and supersession, forget barriers, and deletion.
- The V1 pipelines: one Remember pipeline, a full-text Recall pipeline, an embedding Recall pipeline, a deterministic exact Forget pipeline, and a small static router that picks one pipeline per operation.
- Derived retrieval indexes (full-text and vector). They are rebuildable projections of canonical Memory and never decide what is true.
- A bounded Agent Source lane: typed source units from one Runtime snapshot, indexed for semantic search under the same per-Agent and per-snapshot isolation. It is derived state, not world truth.
Long-term Knowledge also belongs to Cognition's domain, but V1 has no active Knowledge path: nothing reads or writes Knowledge yet, and Memory never promotes itself into Knowledge.
What Stays With Runtime And Realm
Runtime keeps running the show: LocalAgent identity, authorization, the Conversation and its committed events, context planning, AI execution, and the final commit of each turn. Authorization is evaluated by Runtime before any Memory operation crosses the Bridge, and Runtime is the only caller Cognition accepts.
Realm keeps world truth. A memory about something that happened in a world does not rewrite that world's state or history.
Apps and SDKs never touch Cognition. They see Memory only as bounded results returned through Runtime.
Reader Scenario: Remembering You Across Sessions
Suppose on Monday you tell your Agent you're allergic to peanuts. That evening you close the app. On Friday you ask it to suggest a snack.
- On Monday, Runtime committed your message as an event and delivered it to Cognition through the Bridge.
- Cognition's Remember pipeline judged it worth keeping and proposed the memory; Cognition Core committed it atomically, with provenance marking it as something you explicitly said.
- On Friday, while planning the turn, Runtime asks Cognition for relevant memories. Recall returns the allergy as a hit.
- Runtime decides how to use that hit in context. The snack suggestions steer clear of peanuts.
If the recall index had still been building on Friday, step 3 would have returned "unavailable" instead of a hit, and the turn would have continued anyway, just without that memory.
Reader Scenario: Correcting And Forgetting
Suppose your Agent believes you work at Acme. You tell it: "I left Acme, I work at Beta now."
- Your correction arrives as a new committed event.
- Cognition makes the corrected fact current and removes "works at Acme" from default recall, keeping it only as provenance-bounded superseded history.
Later you go further: "Forget everything about my old job." An explicit forget commits a barrier against those memories. From then on, no restart, retry, index rebuild, or replay can revive them. If you mention your old job again someday, that becomes a new memory formed from a new event, not the forgotten one coming back.
Because banks are Agent-private, none of this touches any other Agent you talk to. Another Agent's memory of you is its own.