Cognee
Type: full-code · Vendor: Topoteretes · Language: Python · License: Apache-2.0 · Status: active · Status in practice: emerging
Knowledge-graph-backed memory control plane that turns raw documents, conversations, and structured data into a queryable graph of entities and relationships, paired with a vector store for semantic similarity.
Description. Cognee is an open-source memory framework that ingests heterogeneous sources (text, code, conversations, structured records) and runs a `.cognify` pipeline that converts plain text into chunks, embeddings, summaries, nodes, and edges. The graph store captures entities and relationships in a knowledge graph, and a parallel vector store holds embeddings for semantic similarity, so agents can search by meaning and by explicit relationships from a single query API. Cognee positions itself as a memory control plane combining embeddings, graphs, and cognitive-science approaches.
Agent loop shape. Pipeline-style ingestion plus query API. The application calls cognee.add() on raw sources and cognee.cognify() to run a six-task extraction pipeline; chunks, embeddings, summaries, nodes, and edges are persisted into a graph store and a parallel vector store. At agent turn time the application calls a search API that can combine vector similarity with graph traversal, and injects results into the prompt.
Primary use cases
- knowledge-graph memory for agents
- hybrid graph + vector retrieval
- long-term semantic memory over heterogeneous sources
- entity-relation extraction for agent context
Key concepts
- .cognify pipeline (docs) — Six ordered tasks that turn ingested plain text into chunks, embeddings, summaries, nodes, and edges in Cognee's vector and graph stores.
- Graph store → knowledge-graph-memory (docs) — Captures entities and relationships in a knowledge graph (nodes and edges that show connections between concepts).
- Vector store → vector-memory (docs) — Holds embeddings for semantic similarity; numerical representations that find conceptually related content.
- Dual storage architecture → semantic-memory (docs) — Vector + graph storage gives both semantic search and structural reasoning over the same data.
- Triplet extraction (docs) — Cognee asks the LLM to turn each chunk into graph nodes and edges as (start_node, relationship_name, end_node) triplets, with names inferred from content.
Patterns this full-code implements —
- ★Knowledge Graph Memory
Cognee's core artefact is a typed knowledge graph built by the `.cognify` pipeline: nodes are entities, edges are relationships, and they are committed to a graph database.
- ★Semantic Memory
Cognee's dual storage architecture exposes both semantic search and structural reasoning; summaries and embeddings are indexed in the vector database for semantic search, while the graph holds entiti…
- ★★Vector Memory
Cognee maintains a parallel vector store of embeddings produced by the `.cognify` pipeline, so hybrid graph+vector retrieval is possible from one query call.
References
Cognee documentation — core concepts
“Dual storage architecture gives you both semantic search and structural reasoning”
“The .cognify pipeline is made of six ordered Tasks. Each task takes the output of the previous one and moves your data closer to becoming a searchable knowledge graph.”
topoteretes/cognee on GitHub
“It combines embeddings, graphs and cognitive science approaches to make your documents both searchable by meaning and connected by relationships as they change and evolve.”
Provenance
- Last analyzed:
- Last updated:
- Verification status: verified