Cognee: Open-Source AI Memory for Agents, Explained
Cognee is an open-source memory engine for AI agents. You give it documents, code, chat histories and other data, and it builds a knowledge graph of the people, things and relationships in them, alongside a vector index for searching by meaning. Agents then recall from that memory instead of starting every session from scratch. The whole API comes down to four verbs: remember, recall, improve and forget. Cognee’s repository has about 33,300 stars on GitHub, it runs on your own laptop with small local models and no API key, and there’s a managed Cognee Cloud for teams that don’t want to host it.
The quick answer: Cognee is one of the most popular open-source options for giving agents long-term memory, with an Apache 2.0 license, a Berlin team that publishes its research and benchmark methods, and a $7.5 million seed round from investors who helped build OpenAI and Facebook AI Research. It suits developers who want to own their memory layer and teams building a shared “company brain” for their agents. A disclosure: we make Carly, an AI assistant with its own memory and workspace, so we compete on the memory side. Everything below comes from Cognee’s site, documentation, GitHub repository and announcements.
Cognee at a glance
Checked against Cognee’s site, pricing, README and funding announcements on November 12, 2026.
| Cognee | |
|---|---|
| What it is | An open-source memory engine that turns data into a knowledge graph agents can recall |
| Core API | remember, recall, improve, forget |
| Open source | Apache 2.0, about 33,300 GitHub stars |
| Works with | Claude Code, Codex, Cursor, OpenClaw, Cline and any MCP client; Python, TypeScript, Rust and REST |
| Hosting | Local, self-hosted, Cognee Cloud, or in your own cloud on Enterprise |
| Funding | $1.5 million pre-seed (2024) and a $7.5 million seed led by Pebblebed (2026) |
| Company | Topoteretes UG, Berlin, with an office in San Francisco; founded by Vasilije Markovic |
What Cognee does
Cognee’s homepage frames the problem in three parts: answers scattered across documents, tickets and chat sessions; agents that solve the same task they already solved last week; and company rules no model was trained on. Its answer to each is the same memory layer, which “connects all your data into one brain.”
The README lists three main uses:
- A company brain. Documentation, conversations, tickets, code and agent work go into shared memory, so people and agents can trace a decision back to the discussion and code behind it.
- Memory across agent runs. Agents keep project context, past decisions, fixes and learned rules, and useful lessons from one session become knowledge the next session can find.
- Agents grounded in your domain. Memory can be shaped around the entities and relationships an application cares about, using custom data models and ontologies.
Cognee’s published case studies include research memory at Bayer and an agentic teaching assistant at the University of Wyoming.
How it works
Cognee combines three kinds of storage in one engine: a relational database for metadata, a vector store for meaning-based search, and a graph for how facts connect. Text becomes entities, relationships and searchable chunks, and code becomes a graph of symbols and dependencies. At query time, Cognee picks graph, vector or code context depending on the question.
The four operations in its API do the following:
| Operation | What it does |
|---|---|
| remember | Stores content or code in permanent memory, or in a session |
| recall | Retrieves context and answers, routing to a search strategy automatically or using one you choose |
| improve | Enriches memory, applies feedback and moves useful session knowledge into the graph |
| forget | Removes a specific item or dataset |
The improve step is where Cognee puts the most emphasis. In the cognee 1.0 announcement, Markovic wrote that “Most memory tools are built to store more; cognee is built to get better with use.” When agents retrieve, correct, reuse or ignore information, those signals feed back into the graph, so a correction made in one session is less likely to be needed again.
Recent releases pushed toward running without a cloud model. Since version 1.6.0 in September 2026, Cognee can build and search text memory using a small local extraction model called GLiNER and a local embedding model, with no LLM key. Adding an OpenAI or other model key turns on generated answers and the improvement stages.
What changed in cognee 1.0
Cognee 1.0 replaced the older pipeline-style API (add, cognify, search) with the four verbs above, added a TypeScript SDK and a Rust engine for lightweight and edge environments, and made it possible to export memory in an open format so it isn’t locked in. The announcement was candid about why: “almost nobody wants to run a graph database.” Teams didn’t want another system to license and secure, and individual developers just wanted something that installs and works.
One result is that the whole memory layer can now run on a single Postgres database instead of a separate graph database, vector database, Redis and relational store. The README is clear that Postgres as a graph store is a demo feature in the open-source version, and the production-ready version is a licensed product.
Connecting agents
Cognee ships plugins for Claude Code, Codex and OpenClaw, and an MCP server for Cursor, Cline and other MCP clients. The coding-agent plugins use hooks to capture prompts, answers and tool activity, store them in graph memory, and recall past decisions in the next session. For applications, there are Python, TypeScript and Rust libraries and a REST API, plus integrations with frameworks like the Claude Agent SDK, OpenAI Agents SDK, LangGraph, Google ADK and n8n.
Teams switching from another memory system can import memory from Mem0, Letta, Zep or Graphiti. Cognee Cloud’s paid plan adds data source connectors for Slack, Notion, Linear and Google Drive, and the SDK added Google Drive and Gmail connectors in September 2026.
Benchmarks, with their caveats
Cognee publishes its evaluation methods, which is less common than it should be in this category. Its headline results are a score of 0.79 on the BEAM benchmark at 100,000 tokens and about 91% of 1,762 questions answered correctly on LoCoMo, a test of long-term conversational memory, using what it describes as a strict judge.
The README adds context that’s worth reading before comparing numbers: the BEAM 100K result comes from four evaluation rounds over 20 questions from one held-out conversation, and Cognee says the two BEAM settings it reports used different conversations and procedures. Cognee has also published a post on how LoCoMo scores get inflated across the industry, from answer keys leaking into ingestion to lenient judges. Its own conclusion on its homepage: “We think most standard memory evals are bad.”
Pricing
Cognee’s open-source engine is free to run on your own machines. Cognee Cloud has three plans:
| Plan | Price | What it includes |
|---|---|---|
| Free | $0 | One workspace, 1 million tokens, unlimited users and API calls, Claude Code, Codex and MCP integrations |
| Standard | $1 per million tokens processed | Additional workspaces at $5 a month each, Slack, Notion, Linear and Google Drive connectors, code indexing |
| Enterprise | Custom | Deployed in your own cloud with Cognee’s engineers, plus time-aware memory with conflict resolution, sources on every answer, per-user personalization and support SLAs |
Enterprise is sold as a fixed engagement of 6, 12 or 24 months, with a discounted startup package for companies before Series B. Cognee Cloud runs on gpt-oss-120b, OpenAI’s open-weight model.
The company
Cognee is made by Topoteretes UG, based in Berlin, with an office in San Francisco. Vasilije Markovic, its founder and CEO, wrote in the seed announcement that he spent more than a decade in big data engineering, then went back to study cognitive science and clinical psychology to understand how human memory works. He and his team started Cognee in Berlin in 2024.
Cognee announced a $1.5 million pre-seed round in November 2024. In February 2026 it raised a $7.5 million seed round led by Pebblebed, the firm run by OpenAI co-founder Pamela Vagata and Facebook AI Research founder Keith Adams, with 42CAP, Vermilion Ventures and angels from Google DeepMind, n8n and Snowplow. At the time, Cognee said its pipeline runs had grown from about 2,000 to more than a million in 2025 and that it was running live in more than 70 companies. It also joined UC Berkeley’s Xcelerator program for agentic AI startups in spring 2026.
We didn’t find independent customer reviews of Cognee on review sites. The clearest outside signal is its open-source adoption: its GitHub stars roughly tripled between the February announcement, when Cognee cited more than 12,000, and November.
How Cognee compares
Cognee competes with other memory layers, including Supermemory, Honcho, Zep, LangMem, Mem0 and Letta. Compared with Zep, which keeps its service closed and open-sources its Graphiti framework, Cognee’s full engine is open source and runs locally without a cloud model, though some production features, like the Postgres graph store and time-aware conflict resolution, sit in paid tiers. Supermemory leans toward a hosted API with user profiles, and Honcho toward reasoning about individual users. Cognee’s emphasis is on structured company knowledge and memory that improves from feedback. Our guide to connecting Mem0 to an AI agent covers another rival.
Cognee is a memory engine for teams building agents. If you want an assistant that already remembers you without building anything, Carly has memory built in: it saves named memory documents that it reads and searches in later conversations, searches a knowledge base of files you upload by meaning, and keeps your to-dos, data tables and CRM in one workspace. It’s ready the day you sign up, works inside ChatGPT or Claude or by text, and also labels and sorts email across all your accounts, handles advanced scheduling, and makes and answers calls and texts.
Who Cognee suits
Cognee is a good fit if:
- You want an open-source memory layer you can run and inspect yourself.
- You’re building a shared memory of documents, tickets, code and agent work for a team.
- You want coding agents like Claude Code or Codex to remember decisions across sessions.
- You need memory shaped around your own domain, with custom ontologies.
Look elsewhere if:
- You want a finished assistant rather than infrastructure.
- You need production graph storage on Postgres or time-aware conflict resolution without a paid plan.
- You want a long record of independent customer reviews before committing.
FAQ
What is Cognee?
Cognee is an open-source AI memory engine. It turns documents, code and conversations into a knowledge graph with vector search, so AI agents can remember and recall information across sessions.
Is Cognee open source?
Yes. Cognee’s engine is licensed Apache 2.0 and has about 33,300 GitHub stars. Some production features, such as the Postgres graph store, are part of paid offerings.
Does Cognee need an OpenAI key?
No. Since September 2026 it can build and search text memory with small local models. An LLM key adds generated answers and the memory improvement stages.
How much does Cognee cost?
The open-source engine is free. Cognee Cloud has a free plan with 1 million tokens, a Standard plan at $1 per million tokens processed, and custom Enterprise engagements.
Who founded Cognee?
Vasilije Markovic founded Cognee, made by Topoteretes UG in Berlin. It raised a $7.5 million seed round led by Pebblebed in February 2026.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
