Cognee 1.0 Launches: Open-Source AI Agent Memory Gets a Cloud, a Rust Core, and Single-Postgres Deployment (2026)

On June 26, 2026, cognee released cognee 1.0, the open-source memory platform for AI agents. The release moves cognee from a developer library to a production system: the same engine now ships as a managed service called Cognee Cloud, runs on a single Postgres database, is rebuilt around a high-performance Rust core, and is reorganized around a memory-native API. For teams building agents that need to remember context across sessions rather than starting over each time, cognee 1.0 is the first version the company positions as ready for production deployment rather than experimentation.

Why do AI agents need persistent memory?

Cognee solves the problem that agents without persistent memory repeat work, re-spend tokens, and lose context between sessions. The common workaround is to stitch together retrieval-augmented generation, a vector store, and application logs, which becomes brittle and expensive at scale. Cognee's approach is structured, governed memory built on a dynamic knowledge graph that connects relationships across everything an agent has learned and improves as the agent uses it. The category cognee is staking out is memory for AI agents, distinct from raw retrieval, and 1.0 is its argument that this layer deserves dedicated infrastructure rather than improvised plumbing.

What's new in cognee 1.0?

The most structural change in cognee 1.0 is a memory-native API. Where earlier versions exposed the steps cognee performed internally — add, cognify, search — 1.0 reorganizes the surface around four verbs agents actually use: remember, recall, improve, and forget. Alongside the API, Cognee Cloud is a new managed SaaS that the company says moves a project from local to production in one line, with no separate infrastructure to operate. A continuously self-improving memory loop rounds out the release, designed to get better with use by carrying forward only what matters as agents retrieve, correct, and reuse information.

The release also rebuilds cognee's foundations and its reach. A new Rust core targets on-device and edge memory in lightweight environments, while a first-class TypeScript SDK extends cognee to Node-based agents alongside the existing Python support. A new open format called COGX lets teams export and import memory to and from JSON, which matters for portability and inspection: memory built in cognee can be moved out rather than trapped in a proprietary store. Together these changes are what let cognee describe 1.0 as production infrastructure rather than a library, deployable from a managed cloud down to the edge.

How does cognee connect to agents like Claude Code and Cursor?

Cognee 1.0 attaches to popular coding and chat agents through a single install command, pip install cognee, plus a configuration file. The company lists support for Claude Desktop, Claude Code, Cursor, Codex, OpenClaw, Windsurf, Gemini CLI, Cline, plain REST, and other MCP-compatible agents. The significance is breadth: rather than locking memory to one assistant, cognee positions itself as a portable layer that any Model Context Protocol agent can adopt with minimal setup. For developers already running several agents across a workflow, a shared memory backend that each tool can read from is a more interesting claim than any single integration.

How does cognee perform on the BEAM memory benchmark?

Cognee reports scoring 79% on the public BEAM memory benchmark at a 100,000-token context window, against a reported state of the art of 73.4%, and 67% at a 10-million-token context window, against 64.1%. It also reports that token usage stays roughly flat as the stored dataset grows, where a long-context approach tends to increase token spend more or less linearly. These are cognee's own reported results, so teams evaluating the platform should review the benchmark methodology directly rather than taking the headline figures at face value.

Cognee's own framing around these numbers is notably measured. The company states that a benchmark is not a definitive score, and that the point is a consistent pattern across multiple tests rather than any single result. That pattern, in cognee's telling, is that structured and persistent memory degrades less as stored context grows than simply enlarging a model's context window does. For a category where the hard problem is staying accurate as memory accumulates, that is the claim that matters, and it is the one independent testing will most need to confirm as the field matures.

Why does running on a single Postgres matter?

Running on a single Postgres matters because it removes the multi-system overhead that usually comes with agent memory. The common setup runs a separate vector database, a graph store, and a relational database side by side, each adding cost and operational complexity. Cognee 1.0 collapses this onto one Postgres instance, with no standalone vector or graph engine to operate. Because Postgres is familiar to most engineering teams and easy to back up and inspect, consolidating onto it lowers the burden of adding memory. The company frames this as a defining decision in 1.0, and it should resonate with smaller teams.

Is cognee genuinely open source?

Cognee's core is open source, and the company emphasizes that users keep full ownership of their data and control over their schema, giving them memory they can inspect and govern; the open COGX export format reinforces that portability. This positioning distinguishes cognee from closed memory services where the data model is opaque. The managed Cognee Cloud offers a hosted path for teams that do not want to run infrastructure, while the open-source core remains available to self-host. That dual model, an inspectable open core plus an optional managed service, is a familiar and credible structure in developer tooling.

Who's using cognee, and who's funding it?

Cognee reports that around 6 million memories are created on the platform each month and that more than 100 companies run on it, including Bayer, where the company says cognee supports scientific research and hypothesis generation. Cognee is seed-stage and backed by Pebblebed, the firm run by Pamela Vagata, part of OpenAI's founding team, and Keith Adams, formerly of Facebook AI Research, alongside 42CAP and Vermilion Cliffs, with angel investors from Google DeepMind, n8n, and Snowplow. The investor roster and named enterprise usage give the launch more weight than a typical seed-stage open-source release.

How does cognee compare to other open-source memory tools?

Against other open-source and hosted memory tools, cognee 1.0's differentiators are specific: a memory-native remember/recall/improve/forget API, a self-improving graph-based memory loop, single-Postgres deployment, a Rust core for lightweight and edge environments, and broad MCP agent support out of the box. Its reported BEAM results and flat token usage as data grows are the headline performance claims, though independent verification will matter as the category matures. For teams currently duct-taping RAG and vector stores together, cognee 1.0 is worth evaluating as a consolidated alternative, particularly where data ownership and operational simplicity are the priorities.

The takeaway

Cognee 1.0 is a credible step from library to platform. The combination of a memory-native API, a managed Cloud, single-Postgres deployment, and one-line agent connections lowers the barrier to giving agents real memory, while the open-source core keeps the data inspectable and portable. The benchmark numbers are strong but are the company's own and deserve hands-on validation before teams rely on them. For developers building agents that should remember rather than reset between sessions, cognee 1.0 is one of the more complete open-source options to land this year, and a clear marker of where agent memory is heading.

Full launch details are available in cognee's official 1.0 announcement. For more, see the cognee documentation, the open-source repository on GitHub, and cognee on LinkedIn and X.

‍