- CollectionMemory & State
- Classificationagent-native
- MCPAvailable
- Agent SkillUnknown
- InterfacesMCP · REST · SDK
What it does
Cognee is a memory control plane that ingests an agent’s data sources (warehouses, vector stores, files, APIs), automatically extracts ontologies, builds and maintains a managed world model (graph + vectors), and serves recall back to agents. The pitch is “model your agent’s world” — instead of stuffing raw chunks into a context window, the agent reasons over a structured representation of what it knows.
Three operational stages:
- Ingest — connect Snowflake, Postgres, Slack, PDFs, MD files, REST APIs.
- Reason — Cognee builds the ontology + managed graph + recall tuning that compounds with use.
- Act — agent calls
cognee.search(...)from Claude Code, LangGraph, OpenClaw, MCP-compatible runtimes, etc.
The library is open-source; Cognee Cloud is the optional managed tier (free locally; pay only when scaling).
Agent-facing interfaces
- Catalog protocols: MCP, REST, SDK
- MCP: ✅ Available — first-party MCP server.
- Agent Skill: ⚠️ No SKILL.md published by Cognee directly, but Cognee is listed as a runtime target by other catalog entries’ skills.
- Onboarding type:
sdk
How agents onboard
pip install cognee
# then:
import cognee
await cognee.add(documents)
await cognee.cognify() # build the world model
results = await cognee.search("...")
See the full onboarding section in the dossier.
Protocol surface
| Interface | Detail |
|---|---|
| Python SDK | pip install cognee |
| 28+ connectors | Snowflake, Postgres, files, REST, Slack, … |
| MCP server | First-party |
| Framework adapters | Claude Code, Codex, Cursor, LangGraph, CrewAI, Continue, Hermes, OpenClaw |
| Cloud + On-Prem tiers | Free local; paid hosted; enterprise on-prem |