Skip to main content
Cognee Cloud turns what you give it into memory that agents and people can ask questions of. This page explains the ideas behind that: where memory lives, what it is made of, and the operations that write, read, improve and remove it. Each section links to the Cloud page that shows how to do it from the app or the HTTP API.

Where memory lives

Each workspace keeps its data, its vector embeddings and its knowledge graph in its own managed PostgreSQL database. Cognee manages the language and embedding models, so there is no model key to bring. See Architecture.

What the graph is made of

When Cognee processes your data, it turns it into DataPoints: typed records such as documents, chunks, entities, entity types and summaries. Each one is stored as a graph node with an embedding, so it can be found both by meaning and by following its connections. Recall searches over all of them, not only the entities the Graph page draws. The entity types come from an ontology. By default Cognee uses generic types such as Person, Organization and Concept. With the Premium auto-ontology, Cognee builds entity types from your own data, and you can refine them in plain language. You can also upload your own ontology file and pass it when you add data. See Configuration & Ontologies. Cloud does not accept custom graph models. To steer what gets extracted, use an ontology or a custom extraction prompt.

The four memory operations

The older operations still work as lower-level steps. Add uploads content without processing it, cognify builds the graph from what was added, and search runs one search type you pick. Use them when you want to upload many files first and build the graph once. remember is add and cognify in one call, and recall picks the search type for you.

Organizing and scoping memory

Datasets are the main boundary. Each one has an owner, the person who created it, and can be shared with teams and people. Permissions decide who, and which agents, can read or write it, and an agent never reaches more than its owner can. See Memory Permissions. Node sets are tags inside a dataset. Pass node_set when you remember or add content, and later pass those names as node_name to limit a recall or search to that part of the graph. See Scoping by node type.

Sessions and self-improvement

A session holds one agent conversation. Session entries are kept until a conversation has had no activity for one year. Sessions become long-term memory through improve. It weights existing memories by the feedback in the session, saves the questions and answers, distills reusable lessons, and adds them to the dataset’s graph, so the next session can recall what this one learned. When you connect a coding agent through Agent Integrations, the plugin runs this for you. In the app it is called Self-improvement. The same run updates personalized ranking. Cognee reads how you rated earlier answers, both ratings you send and ones it infers from what you say next, and nudges your own future results toward what helped you. Your ratings never change a teammate’s results. See Personalized ranking.

Moving memory

A dataset’s knowledge graph can be exported as a COGX archive (.cogx.tar.gz), a portable file that another Cognee instance can import. See COGX archive and Syncing a local instance.

Going deeper

Cognee Cloud runs the open-source Cognee engine. The Cognee Core Concepts pages explain that engine in more depth, with examples for the Python library. Their code and server settings are for self-hosted Cognee and don’t apply to your Cloud workspace.

DataPoints

The typed records the graph is built from.

Remember

How content becomes graph-backed memory.

Recall

How recall picks a search strategy.

COGX

The format behind dataset export and import.