Cognee
Also known as: topoteretes cognee
Open source AI memory engine combining relational, vector and graph stores, with remember, recall, improve and forget operations, session memory, and Python, REST, CLI and MCP access.
Cognee, from Topoteretes UG in Berlin, is an open source AI memory engine that turns documents, code and application data into persistent memory that agents and applications can store, query and improve over time. It combines three stores. A relational store holds documents, chunks and their provenance, a vector store handles semantic similarity, and a knowledge graph holds entities and relationships. Pipelines that ingest, extract and enrich data build them, with ontologies, a context index for each dataset and deduplication to keep the graph coherent.
The developer surface centers on four operations, remember, recall, improve and forget. Sessions give each user short term memory for a conversation or agent run, session distillation turns guidance from a session into permanent lessons, and preferences for each user adjust retrieval. Recall routes across vector, graph and hybrid retrieval with provenance on answers. Cognee is reached through a Python API, a REST API, a CLI and an MCP server, integrates with LangGraph, CrewAI, the Claude Agent SDK, the OpenAI Agents SDK, Google ADK, n8n, Claude Code and Codex, and ingests from Slack, Notion, Linear, Google Drive and GitHub on paid plans.
Customers choose their LLM, embedding provider, vector store and graph store, and run Cognee locally, self hosted with Docker or Kubernetes, in Cognee Cloud, or in their own cloud through an Enterprise engagement. Multi user mode adds tenants, roles and access control lists, with a separate database for each tenant. Cognee states it does not yet hold SOC 2 or ISO 27001. The cloud has a free tier with one workspace and a million tokens, a Standard plan at $1.00 per million tokens plus $5 per extra workspace, and Enterprise engagements that bring your own cloud.
Vendor details
Canonical URL
https://www.cognee.ai
Category
Agent infrastructure
Subcategory
Memory
Funding status
Independent. Raised a $7.5 million seed round in early 2026. Open source (Apache 2.0) with more than twelve thousand GitHub stars and 80 plus contributors. Runs live in more than seventy companies including Bayer and the University of Wyoming, processing over a million pipeline runs a month.
Company status
independent
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
Memory native API (remember, recall, forget, improve) through a Python API, a REST API, a CLI and an MCP server. Native integrations with LangGraph, CrewAI, the Claude Agent SDK, the OpenAI Agents SDK, Google ADK, n8n, Claude Code and Codex, and data source integrations for Slack, Notion, Linear, Google Drive and GitHub. Customers choose their own LLM, embedding provider, vector store and graph store.
In practice
Your agent forgets everything when a session ends. You add Cognee's remember and recall API, and it builds a persistent knowledge graph so the agent carries context, preferences, and decisions across sessions.
You need multi hop reasoning over connected documents, not flat chunks. Cognee's pipelines extract entities and relationships into a knowledge graph, so retrieval traverses relationships instead of nearest neighbor lookups alone.
Regulated data cannot leave your infrastructure. You self host the open source engine on your own graph and vector stores, keep full data ownership, and still give every MCP agent one shared memory.
Sources & related URLs
Related / legacy domains
Agentic Index coverage score
7.0 / 14 capabilities · 50%
| Integrations & Tool Calling | Partial |
|---|---|
|
The Standard plan adds OAuth data source integrations for Slack, Notion, Linear and Google Drive, and a GitHub integration reindexes repositories on each push. All of them read content into memory. Native integrations wire Cognee into LangGraph, CrewAI, the Claude Agent SDK, the OpenAI Agents SDK, Google ADK, n8n, Claude Code and Codex. Cognee names no integration that lets an agent write to or act in outside systems. SourceCognee, cognee.ai/pricing and docs.cognee.ai integrationsread 2026-09-21 |
|
| Workflow Orchestration | Partial |
|
Pipelines orchestrate Tasks into coordinated data processing workflows with a typed runtime context, and developers can build custom tasks and pipelines. They sequence the steps that build and enrich memory. Cognee names no control flow for the customer's own agent workflows. SourceCognee, docs.cognee.ai core-concepts/building-blocks/pipelines and guides/custom-tasks-pipelinesread 2026-09-21 |
|
| Knowledge Grounding & RAG | Full |
|
Documents, code and application data, including Slack, Notion, Linear, Google Drive and GitHub sources, are ingested into three stores. A relational store holds documents, chunks and provenance, a vector store holds embeddings, and a knowledge graph holds entities and relationships. Recall queries them by vector, graph or hybrid retrieval, with provenance on answers and an optional context index for each dataset. The index and graph over the customer's knowledge persist and stay queryable. SourceCognee, docs.cognee.ai core-concepts/architecture and cognee.ai/pricingread 2026-09-21 |
|
| Human Oversight & Guardrails | Not documented |
|
A feedback system lets users rate answers to improve memory, and the permission system controls who can read and write datasets. Cognee names no approval step, consent checkpoint, escalation rule or pause for a person to sign off before an agent acts. SourceCognee, docs.cognee.ai guides/feedback-system and full documentation exportread 2026-09-21 |
|
| Security, Identity & Governance | Partial |
|
In multi user mode, Cognee enforces authentication and permissions scoped to each dataset through users, tenants, roles and access control lists, and each tenant's data sits in a dedicated database. Larger enterprise engagements add SSO. Cognee says it does not currently hold SOC 2, ISO 27001 or an equivalent third party certification. SourceCognee, docs.cognee.ai cognee-cloud/functionality/data-and-security and core-concepts/multi-user-mode/permissions-system/overviewread 2026-09-21 |
|
| Observability & Auditability | Partial |
|
Agent session traces record what a decorated agent function did on each call, including calls that raised errors, and can be recalled later. Cognee emits OpenTelemetry spans and logs to an OTLP endpoint such as Langfuse, writes structured logs and visualizes memory provenance. The traces cover Cognee's own operations and the decorated call. Cognee names no step by step view of the agent's prompts, tool calls, retrieved knowledge and outputs. SourceCognee, docs.cognee.ai guides/agent-session-traces, setup-configuration/logging and integrations/opentelemetry-tracingread 2026-09-21 |
|
| Memory & State Persistence | Full |
|
Sessions are short term memory scoped to a user and session ID, holding recent queries, responses and the context used to answer them. Session distillation turns guidance stated in a session into permanent lessons that later recalls respect. User preferences adjust retrieval for each user and agent. Forget can reset only the memory state or delete one dataset without touching others. SourceCognee, docs.cognee.ai core-concepts/sessions-and-caching, guides/session-distillation and core-concepts/main-operations/forgetread 2026-09-21 |
|
| Deployment & Data Residency | Full |
|
Customers can use managed Cognee Cloud or self host the open source software with Docker Compose, a Helm chart for Kubernetes, EC2 or Coolify. Cognee also runs fully locally with Ollama and embedded stores, and Enterprise customers can deploy it in their own cloud. SourceCognee, docs.cognee.ai how-to-guides/cognee-sdk/deployment and cognee.ai/pricingread 2026-09-21 |
|
| Prebuilt Agents, Templates & Packs | Not documented |
|
Examples and guides show how to build agents on Cognee's memory layer. Cognee names no catalog or set of ready made agents or templates for a customer to adopt. SourceCognee, cognee.ai/pricing and docs.cognee.ai examples/overviewread 2026-09-21 |
|
| Triggers & Channel Coverage | Not documented |
|
Agents and applications call Cognee, a memory layer, through remember and recall. Cognee names no schedule, event or channel that starts an agent's work without a person or the calling application starting it. SourceCognee, docs.cognee.ai (full documentation export)read 2026-09-21 |
|
| Model Flexibility & Routing | Full |
|
The customer configures which LLM provider Cognee uses, from OpenAI, Azure OpenAI and Anthropic to local models through Ollama, and separately chooses the embedding provider, vector store and graph store. SourceCognee, docs.cognee.ai setup-configuration/llm-providers and embedding-providersread 2026-09-21 |
|
| APIs, SDKs & MCP Extensibility | Full |
|
A REST API with an OpenAPI specification covers datasets, ingestion, recall and activity. Cognee also offers a Python API, a CLI and an MCP server for Cursor, Claude Code and other clients, and serve and push connect the SDK to Cognee Cloud or a remote instance. SourceCognee, docs.cognee.ai api-reference/introduction, python-api and cognee-mcp/mcp-overviewread 2026-09-21 |
|
| Testing, Debugging & Optimization | Not documented |
|
Memory benchmarks published by Cognee measure its own engine. Cognee names no evaluation harness, scored test cases or quality gate a customer runs against its own agent. SourceCognee, cognee.ai/pricing and docs.cognee.ai (full documentation export)read 2026-09-21 |
|
| Browser & Computer Use | Not documented |
|
Web URL ingestion fetches page content with custom extraction rules to build memory, so it builds the corpus and no agent operates a browser at run time. Cognee names no browser, desktop or computer control. SourceCognee, docs.cognee.ai guides/web-url-ingestionread 2026-09-21 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Product updates
Version 1.6.8 of Cognee replaces the Slack history importer with a Slack connector that syncs owner selected channels into memory, carrying edits and deletions through. It adds an @cognee Slack app that answers a question in its thread from what channel members can read, answers direct messages from the member's channels and datasets, and offers Remember this and Forget this message shortcuts.
Bears on: Agent capability
View sourceFacts that contradict each other in a Cognee dataset can now be listed and settled through the API and SDK in v1.6.6, by picking one fact with an optional reason. The decision is recorded and kept by later automatic reviews, and retrieval uses the chosen fact while the rejected ones stay stored. The improve step now also reviews conflicting facts by default.
Bears on: Memory / state
View sourceRelease 1.6.5 of Cognee adds an opt in fact conflict review to improve(), replacing its earlier contradiction detection, and a BROAD search type for questions about a whole dataset. The release also adds a Google Calendar connector, turns raw Cypher search off by default and deprecates memify in favor of improve().
Bears on: Agent capability
View sourcePricing
Free (1M tokens) · Standard $1.00 per 1M tokens · Enterprise BYOC · open source
Tokens processed, plus a fixed fee per additional workspace
Included quota
Free ($0): 1 workspace, 1M tokens, unlimited users and API calls, agentic integrations (Claude Code, Codex, MCP). Standard: $1.00 per 1M tokens processed plus $5 per additional workspace a month; data source integrations (Slack, Notion, Linear, Google Drive), code indexing, in-app support. Enterprise: a fixed-scope bring-your-own-cloud engagement (Startup 12 months, or 6, 12 or 24 months) with dedicated support and SLAs.
What is public
The free tier, the Standard rate of $1.00 per million tokens plus $5 per extra workspace, the Enterprise engagement tiers and the free open-source self-host path are published.
Billing mechanics
Open source and free cloud tiers carry the engine at no vendor cost. Cloud Pro meters tokens processed at $2.50 per million. Enterprise runs in a private cloud with bring your own key so model and infrastructure costs stay on the customer's accounts.
Cost watchouts
Building memory spends LLM tokens at ingestion, billed per million tokens on Standard, and each extra workspace adds $5 a month; self-hosting moves the cost to the customer's own LLM keys and stores.
Variable cost rationale
Token based cloud pricing scales with ingestion and query volume, but the free tier and modest $2.50 per million rate keep growth predictable, and self hosting caps the vendor cost entirely.
Overage / add-ons
Cloud usage bills at $2.50 per 1 million tokens processed above the free allowance.
Sales call required
Mixed (some tiers require a call)
Free / trial
Free forever: 1 workspace and 1M tokens included, no card required; open source free to self host
Lowest paid plan
Standard at $1.00 per 1M tokens, plus $5 per additional workspace
Commercial notes
Open source under Apache 2.0 with a free cloud tier and usage-based Standard plan; Enterprise is delivered as a fixed-scope deployment engagement in the customer's own cloud.
Key ambiguities
Enterprise engagement prices are not published.
Missing data
Enterprise engagement prices are not published.
Related vendors
- AgentOps — Agent observability and debugging platform: open source SDKs trace…
- AgentPhone — AgentPhone gives the AI agents a team builds their own US or…
- Agno — Python agent framework and AgentOS runtime (formerly Phidata) for…
- AIsa — Resource and payment gateway for AI agents, with one key to 110+…
- Albato — Albato is a no code integration platform that sells an embedded,…
- AlphaBitCore — AI control plane for regulated financial firms: one gateway enforces…
Alternatives to Cognee
The closest documented capability profiles to Cognee among agent infrastructure platforms tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- Zep7.0 / 14Fuller documented coverage on Security, Identity & GovernanceCognee vs Zep →
- Hyperspell7.0 / 14Adds documented Testing, Debugging & Optimization
- Mem06.0 / 14Fuller documented coverage on Security, Identity & GovernanceCognee vs Mem0 →
- Jina AI5.0 / 14Adds documented Browser & Computer Use
- Lemony7.0 / 14Adds documented Human Oversight & Guardrails and Prebuilt Agents, Templates & Packs, among others
- Vercel Sandbox6.0 / 14Adds documented Prebuilt Agents, Templates & Packs
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded