What it is
Cognee builds long-term memory for AI agents. It extracts entities and relationships from documents, code, tickets and conversations into a knowledge graph backed by vector search, and exposes four operations: remember, recall, improve and forget. It can be used from Python, a CLI, REST, MCP, TypeScript and Rust SDKs, and Claude Code and Codex plugins.
The graph, vector and relational stores can be swapped independently. Since version 1.0 the whole memory layer can run on one Postgres instance with pgvector, though Postgres as the graph store is labelled a demo. Since 1.6.0 (18 September 2026), text memory can be built with a local GLiNER extraction model and no cloud API key. Version 1.6.1 added Google Drive and Gmail connectors, and it can import memories from Mem0, Letta, Zep or Graphiti.
Cognee Cloud is the managed service, priced per token processed; an enterprise edition is delivered as a bring-your-own-cloud engagement.
Our review
Cognee aims at a different job from Graphiti or Mem0: one shared, queryable knowledge graph built from many sources and used by several agents. Its swappable storage layers are a genuine advantage, and the keyless local mode suits privacy-sensitive deployments. The repo has about 32,700 stars and ships fast, with six releases between v1.6.0 on 18 September and v1.6.5 on 16 October 2026, and a 1.6.6 release candidate on 23 October.
Security needs attention before deployment. A critical advisory published in September 2026 covers versions 0.3.0 to 1.1.3, where a notebook endpoint passed request bodies to Python's exec(); before 1.1.0 it needed no authentication, so any network client could run code on the server. Versions from 1.2.0 are fixed. Separately, product telemetry is on by default: the .env template ships TELEMETRY_DISABLED commented out, and events carry anonymous and persistent install IDs, with dataset names hashed. Search history logging is also on unless you turn it off.
The API server defaults to multi-tenant mode with authentication, which is the right default. The README's quick docker run example turns access control off for single-user use, and copying that into a shared deployment would expose the API on port 8000.
The maturity gaps are familiar. The 1.0 API replaced the older add, cognify and search calls, so many tutorials are out of date. The bundled GLiNER extractor and the Postgres graph store are both described as demos, with production versions in a licensed product. Cognee reports 0.79 on the BEAM conversational-memory test at 100,000 tokens and 0.67 at 10 million, but warns the two runs used different setups. Its pricing page shows heyData GDPR and EU AI Act badges but no SOC 2, ISO 27001 or HIPAA certification.
For memory of one changing conversation, Graphiti's temporal model is stronger, as Cognee's own comparison concedes. Cognee's case rests on breadth of sources and a migration path in from rivals, which few memory layers offer.
Where it earns its place
A platform team at a 300-person software company
Build one knowledge base from docs, tickets and code that several internal agents query
Self-host the API with Docker Compose, keep authentication on and set TELEMETRY_DISABLED. Postgres with pgvector can hold everything, but the Postgres graph store is a demo, so plan for a dedicated graph backend or the licensed version before production.
A solo developer using Claude Code across several projects
Keep decisions and fixes remembered between sessions
Install the cognee-memory plugin from the topoteretes/cognee-integrations marketplace and run Cognee locally with the GLiNER extra, no API key needed. Free; answers are retrieved passages unless an LLM key is configured, which defaults to OpenAI.
A startup CTO who does not want to run graph infrastructure
Give a customer-facing agent hosted memory
Cognee Cloud is free to 1M tokens, then $1 per 1M tokens processed plus $5 per extra workspace a month, so ingesting 100M tokens a month costs about $100. No SOC 2 report is published, which may block enterprise customers.
Before you commit
- Setup
- uv pip install "cognee[gliner]" on Python 3.10 to 3.14, then remember and recall from Python or cognee-cli; local models download on first use. Docker: the cognee/cognee image, or docker compose with UI and MCP profiles on ports 8000, 3000 and 8001.
- Integrations
- Python, CLI, REST, MCP server, TypeScript SDK, Rust SDK, Claude Code and Codex plugins, an OpenClaw plugin, Google Drive and Gmail connectors, and imports from Mem0, Letta, Zep and Graphiti. Cloud adds Slack, Notion, Linear and Google Drive sources.
- Data and privacy
- Self-hosted data stays in your stores, but telemetry with install IDs is on by default (set TELEMETRY_DISABLED=1) and search history is logged unless disabled. With an LLM key set, documents go to OpenAI by default. Critical advisory GHSA-8pr4-p4c8-8gr9 affects 0.3.0 to 1.1.3; upgrade to 1.2.0 or later. Apache 2.0.
- Cost at scale
- Self-hosting is free; text memory runs on CPU with local models, and LLM provider charges apply once a key is set. Cloud Standard is $1 per 1M tokens processed plus $5 per extra workspace a month; Enterprise BYOC pricing is not published.
- Limits
- About 1,000 open issues and pull requests; weekly releases; demo-grade local extractor and Postgres graph store; temporal reasoning weaker than Graphiti; no published SOC 2 or HIPAA certification.
What's good
- Swappable graph, vector and relational stores
- Keyless local mode for text memory since 1.6.0
- Python, CLI, REST, MCP and agent plugins
- Imports memories from Mem0, Letta, Zep and Graphiti
What's not
- Telemetry and search logging on by default
- Critical RCE advisory for versions before 1.2.0
- Local extractor and Postgres graph store are demos
- Self-reported benchmarks with mixed setups
Best for
Teams building several agents that need a shared knowledge graph assembled from many sources.
Alternatives
Cognee: common questions
- Is Cognee worth it?
- Cognee builds a shared knowledge graph from almost any source, but its defaults need an owner before they meet production. We rate it 3.8 out of 5.
- How much does Cognee cost?
- Open source under Apache 2.0. Cognee Cloud: free up to 1M tokens and one workspace, then $1 per 1M tokens plus $5 per extra workspace a month; Enterprise is a quoted BYOC engagement.
- What are the downsides of Cognee?
- Telemetry and search logging on by default; critical RCE advisory for versions before 1.2.0; local extractor and Postgres graph store are demos; self-reported benchmarks with mixed setups.
- Who is Cognee best for?
- Teams building several agents that need a shared knowledge graph assembled from many sources.
- How much does Cognee cost for a team or in production?
- Self-hosting is free; text memory runs on CPU with local models, and LLM provider charges apply once a key is set. Cloud Standard is $1 per 1M tokens processed plus $5 per extra workspace a month; Enterprise BYOC pricing is not published.
- Is Cognee safe for company data?
- Self-hosted data stays in your stores, but telemetry with install IDs is on by default (set TELEMETRY_DISABLED=1) and search history is logged unless disabled. With an LLM key set, documents go to OpenAI by default. Critical advisory GHSA-8pr4-p4c8-8gr9 affects 0.3.0 to 1.1.3; upgrade to 1.2.0 or later. Apache 2.0.
- What are the best Cognee alternatives?
- The closest alternatives we have reviewed are Graphiti, Mem0, LightRAG and GraphRAG.
Answers checked Oct 23, 2026.