Reference
Memory model
A few ideas explain everything Kognite does: memories live in a scope, each has a kind and a confidence, retrieval is hybrid, and related facts are linked in a graph.
Scopes
A scope is an isolated memory space. Every key is bound to exactly one scope, and a request can only read and write within it — that’s the isolation boundary. There are three kinds:
| Scope | Shared by | Use it for |
|---|---|---|
personal | Just you, across all your tools | Your preferences, your projects, your working context. |
team | Everyone on the team | Shared conventions, decisions, and institutional knowledge. |
agent | One named agent | A product agent’s own memory, isolated from yours. |
The same fact learned in Cursor and recalled in Claude Code works precisely because both point at a key for the same scope. Available scope counts depend on your plan.
Kinds
Every memory has a kind — a lightweight type you can filter on at search time (memory_search accepts a kind) or set at write time (memory_add / POST /api/memories, default fact). Common kinds:
fact— an objective statement (“the DB is Postgres 16”).preference— how the user likes things done.decision— a choice made, with its rationale.episode— the raw turns of a conversation stored byepisode_log.
Confidence
Retrieval returns a confidence score in 0–1 for how strongly a memory matches your query. Use min_confidence to drop weak matches — a good default for automated pipelines that act on recall without a human in the loop.
Retrieval — how search actually works
Kognite doesn’t pick one retrieval strategy; it fuses several, which is why recall holds up on long histories:
- Full-text — Postgres keyword search catches exact terms, names, and IDs.
- Semantic — a vector embedding of the query matches on meaning, not just words.
- Graph expansion — entities in the top hits pull in related memories a plain search would miss.
- Fusion — the candidate sets are combined with reciprocal-rank fusion and returned top-K.
You don’t configure any of this — memory_search and the REST search endpoint do it for you. See the live numbers on the benchmarks page.
The knowledge graph
As memories are written, Kognite extracts entities (people, projects, tools) and the relationships between them into a graph. That graph powers the expansion step above and can be queried directly with the graph_query tool. It’s what lets a search for “the deploy setup” also surface the box, the database, and the person who owns it.
Operations & limits
Every tool call and API request that touches memory is one memory operation, metered against your plan’s monthly quota. Hitting the cap returns 402 until the month resets or you upgrade. Current per-plan limits are on the pricing page.