Memory model
Cognimemo is a note-taker for agents: your agent does the work and tells Cognimemo what it learned; next time, Cognimemo hands that knowledge back. Cognimemo owns durable, typed storage and fast typed recall — it never executes or drives your agent.
Hierarchy
Account → Project → Space (org) → Bank (person/agent) → Memories| Level | What it is |
|---|---|
| Account | You — signed in at app.cognimemo.com. |
| Project | An isolated tenant: its own data, API keys, members, and billing. |
| Space | An organization inside a project. Every space has an implicit org bank. |
| Bank | One subject's memory store (a user, an agent, a session). |
| Memory | A single stored unit — a fact, an event, or a typed block. |
Isolation is strict: an API key for one project can never read another project's data, and banks don't leak across each other.
The three operations
- Retain — store a memory. Embeddings (and, with a real LLM, entities and links) are computed on ingest. See Retain.
- Recall — retrieve memories with a hybrid of semantic, keyword, graph, and temporal search, fused and reranked. See Recall.
- Reflect — generate a synthesized, disposition-aware answer from memories and consolidated observations. See Reflect.
Fact types
Every memory has a type. Three are produced automatically; five more are typed blocks you write directly:
world · experience · observation · procedure · reasoning · preference · correction · profile
Recall can filter to any subset via types, and default recall (no filter) returns them all — so adopting typed blocks never changes existing behaviour.
Lifecycle
Memories decay through states — New → Active → Expiring → Forgotten — so stale knowledge fades while frequently-used memories stay strong. Corrections and explicit updates reset the relevant memories.