Cognee

Cognee is an open-source AI memory platform built by Topoteretes UG, headquartered in Berlin, that transforms raw data into structured knowledge graphs for AI agents. Most memory systems store text chunks in a vector database. Cognee runs an ECL pipeline (Extract, Cognify, Load) that extracts entities, maps relationships, and builds a queryable knowledge graph with embeddings. This gives agents temporal awareness, entity relationships, and feedback loops that pure vector retrieval cannot provide. The core operations are four: remember (store to graph), recall (query with auto-routing), forget (delete), and improve (refine through feedback). When an agent rates a response, that feedback updates edge weights in the graph, so the memory gets sharper with use rather than staying static. Cognee unifies three storage layers (relational, vector, and graph) into a single engine using SQLite, LanceDB, and Kuzu locally, with managed cloud options for production scale. Cognee ships three SDKs. The Python SDK is the most fully documented. A TypeScript SDK, @cognee/cognee-ts, provides Node.js bindings, and a Rust SDK, cognee-rs, is documented across architecture, configuration, guides, an HTTP server and operations. Native integrations include the Claude Agent SDK, the OpenAI Agents SDK, LangGraph, Google ADK, n8n, Amazon Neptune and Neo4j. Cognee is also available as a Claude Code plugin. Cognee publishes its own MCP server. The vendor documents it for Claude, Cursor, Cline, Continue and Roo Code, and the documentation describes both a standalone architecture mode for individual developers and an API mode for teams sharing one knowledge graph across multiple AI clients. Agentic integrations including MCP are available on the free tier rather than gated behind a paid plan. Pricing is usage-based and was restructured to a token model. Free is $0 per month and includes 1M tokens, one workspace, unlimited users, unlimited API calls and agentic integrations, with no card required. Standard is $1.00 per 1M tokens on a pay-for-what-you-process basis, and adds unlimited workspaces at $5 each per month, data source integrations for Slack, Notion, Linear and Google Drive, code indexing for your repositories, and in-app support. Enterprise is quoted rather than published and is delivered as a fixed-scope BYOC engagement in the buyer's own cloud, offered in six, twelve and twenty-four month lengths with a discounted Startup package for pre-Series B companies, and it adds a dedicated Slack channel, a dedicated support engineer and a support SLA. Self-hosted deployment remains free under the open-source license. What the entry price does not buy: the $5 per workspace charge is an add-on for additional workspaces rather than a plan fee, and it stacks on top of token consumption. The free tier is capped at one workspace and 1M tokens. Bi-temporal memory and conflict resolution, provenance on every answer, and personalization per user and agent are Enterprise features rather than Standard ones. Benchmark results on this row are the vendor's own. Cognee reports a score of 0.79 on BEAM 100K, a long-horizon memory benchmark where the evidence needed for an answer is spread across many turns, and reports human-like correctness rising from 0.738 to 0.925 when its graph-completion retriever runs with chain-of-thought. These are vendor-run evaluations rather than third-party audits, and Cognee has not published a LongMemEval result. Best fit: AI engineers who need agents with persistent, structured memory that improves from real usage rather than replaying conversation history, and teams that want graph-based memory without gating it behind an enterprise plan. Who Cognee is not for: teams that want a fully managed, turn-key hosted memory service with minimal configuration, and teams whose procurement requires third-party or LongMemEval benchmark results rather than the vendor's own. Alternatives worth evaluating are Mem0 for a larger community, Zep for structured user memory with session history, and a general-purpose vector database where graph relationship mapping is not needed. Current state Q4 2026: Cognee 1.0 shipped on 26 June 2026, built around a remember, recall, improve and forget API, and the company joined UC Berkeley RDI's Xcelerator 2026 agentic AI cohort in July. The Standard rate was reduced from $2.50 to $1.00 per 1M tokens, and Cognee's newsletter of 24 September 2026 confirmed the new rate, saying it came down with a rewrite of the pricing page. The same newsletter announced graph extraction that runs without an LLM, using the GLiNER2 model for entity and relation extraction. Cognee raised a $7.5M seed round and reports surpassing 5 million SDK runs per month, and publishes Bayer, the University of Wyoming and Knowunity as named customer case studies. The cloud runs on GPT-OSS-120B, OpenAI's open-weight model. Certifications are limited to GDPR and CCPA, stated on the vendor's trust page and privacy policy respectively, with no SOC 2 or ISO 27001 published. Cognee publishes General Terms and Conditions in separate US and EU versions, effective 27 March 2026, and neither those documents nor the privacy policy contains any statement about training on customer data.