An open-source memory platform for AI agents, built by Germany-based Topoteretes. It addresses a core weakness of generative AI ─ context disappears the moment you close the chat ─ by building a knowledge graph from ingested documents and conversations so the same knowledge can be recalled in the next session. Its defining trait is combining graph, vector, and relational retrieval to surface “related memories,” and it connects directly to agents such as Claude Code and Codex through its MCP server. Licensed under Apache-2.0, it has gathered more than 30,000 stars on GitHub. You can start self-hosted on your own machine or server and move to the managed Cognee Cloud as you scale.
Key Features
- Memory structured as a knowledge graph: Ingested data is not kept as loose text fragments. It is structured in stages ─ chunks, extracted entities, concepts, and induced ontologies. Because it retains who relates to what and how, it can answer questions whose wording does not match the source text
- Four operations: remember / recall / forget / improve: Operations are organized around four verbs ─ store, recall, delete, and self-tune memory. Recall routes to a search strategy automatically, so you do not have to choose between graph and vector search yourself
- Connects to existing agents via an MCP server: The official MCP server supports stdio, SSE, and HTTP transports. A Claude Code plugin is also available, letting you retrofit persistent memory onto an agent
- Both a Python SDK and a CLI: Install with
pip install cognee, set your LLM API key, and it runs. You can also work from the command line withcognee-cli rememberandcognee-cli recall - Swappable databases by design: Graph stores include Kuzu (default), Neo4j, PostgreSQL, and Turso; vector stores include LanceDB (default), PGVector, Qdrant, Chroma, Weaviate, and Milvus. Pick what fits your existing stack
- Self-hosted through cloud: You can keep everything local so your data never leaves your environment, or share workspaces on Cognee Cloud. From the Standard plan onward, Slack, Notion, and Google Drive can be connected as data sources
Pricing
| Plan | Price | What’s included |
|---|---|---|
| Free | $0 | 1 workspace, 1M tokens included, unlimited users, unlimited API calls, Claude Code / Codex / MCP integrations |
| Standard | $2.50 per 1M tokens + $5 per additional workspace | Everything in Free, Slack / Notion / Google Drive integrations, in-app support |
| Enterprise | Contact sales | Everything in Standard, dedicated Slack channel, dedicated support engineer, BYO cloud, support SLA |
Pricing is current as of August 2026. No annual plan is listed, and Standard is usage-based. Check the official site for the latest pricing.
Note that self-hosting the open-source edition carries no fee for Cognee itself. The API costs for the LLM and embedding models it uses internally are separate.
Pros & Cons
✅ Pros
- Apache-2.0 open source, so self-hosting keeps your data under your own control
- Combining a knowledge graph with vector search surfaces related information that plain similarity search misses
- The MCP server lets you add memory to existing agents such as Claude Code without major rework
- Graph DB, vector DB, and LLM are each swappable, making it easy to fit into existing infrastructure
- The free tier includes 1M tokens with no limit on users or API calls, so it is easy to try small
⚠️ Cons
- It is a developer tool that assumes Python and database knowledge ─ not something a non-engineer uses on its own
- Self-hosting means deciding your graph DB, vector DB, and LLM key setup yourself, which takes initial effort
- Ingestion (graph building) calls the LLM, so API costs grow with the volume of data you load
- The Standard plan is usage-based, making monthly costs hard to forecast until your usage settles
- The field is still young, and surrounding best practices and case studies are not yet settled
Comparison with Similar Services
| Criteria | Cognee | Mem0 | Zep | LlamaIndex |
|---|---|---|---|---|
| Positioning | Memory platform for agents | Memory layer for agents | Memory service for agents | Data framework for LLM apps |
| How memory is held | Knowledge graph + vector + relational | Vector-centric (graph features offered) | Knowledge graph with a temporal axis | Index / vector store centric |
| License | Apache-2.0 (open source) | Open source + cloud | Open source + cloud | Open source + cloud |
| Deployment | Self-hosted / Cognee Cloud | Self-hosted / cloud | Self-hosted / cloud | Self-hosted / cloud |
| Strengths | Recall that follows relationships, swappable DBs | Ease of adoption | Chronological handling of conversation history | Retrieval and RAG pipelines in general |
Tools for handling “memory” are multiplying, but Cognee stands out by putting the graph front and center as its retrieval mechanism and by letting you swap nearly every component.
Who Is It For
- Developers who want AI agents such as Claude Code to carry memory across sessions
- Teams that want internal documents and meeting notes retrievable with their relationships intact, not just through full-text search
- Organizations that would rather not hand customer data or design documents to an external service and want to stay self-hosted
- Engineers whose RAG accuracy has plateaued and who want to try graph-based retrieval
- Developers who want to add a memory layer while keeping existing PostgreSQL or Neo4j assets
Summary
Cognee is an open-source memory platform that answers the “forgets when the session ends” constraint of AI agents with a knowledge graph. It is Apache-2.0 and self-hostable, and its MCP server plugs into existing agents right away ─ though working with it assumes Python and database knowledge. A realistic path is to ingest a small dataset locally on the free tier, judge the quality of recall, and only then consider Cognee Cloud or a full rollout.