cognee: open-source AI memory platform for agents

June 30, 2026·AI Newsroom·12 min read·11 sources
cogneetopoteretesai-memoryagent-memoryknowledge-graphgraph-rag+21
GitHub repository landing page for topoteretes/cognee, an Apache-2.0 open-source AI memory platform for agents, showing 25.7k stars, 2.4k forks, 207 open issues, 224 open pull requests, 8,426 commits, and a topics list that includes ai-memory, agent-memory, cognitive-memory, context-engineering, and knowledge-graph.
Screenshot: GitHub repository landing page for topoteretes/cognee (github.com/topoteretes/cognee), captured 2026-06-30 via Playwright Chromium via the project's scripts/browser.mjs. License: Apache-2.0 (project) / screenshot used editorially for the cognee article.

The open-source project cognee (Apache-2.0) is the leading self-hosted answer to a problem every team running AI agents at scale hits: the agent has no memory of what it learned last week, last month, or last release. As of 2026-06-30, the topoteretes/cognee repository sits at 25,747 stars, 2,371 forks, 8,426 commits on main, 207 open issues, 224 open pull requests, 121 releases, and an Apache-2.0 license (per the GitHub REST API, snapshot 2026-06-30). The latest release, v1.2.2 “Truth Subspace & Retrieval Improvements” (release notes, 2026-06-26), adds a new retrieval layer on top of the single-Postgres architecture that shipped in cognee 1.0. The project is also the rare open-source agent tool that has a published research paper behind it: Optimizing the Interface Between Knowledge Graphs and LLMs for Complex Reasoning by Markovic, Obradovic, Hajdu, and Pavlovic, submitted to arXiv on 2025-05-30.

What happened

The four-method API is the surface most builders touch first. Cognee exposes four async operations — remember, recall, forget, and improve — for the entire memory lifecycle (README). remember ingests into the permanent knowledge graph (running add + cognify + improve) or a fast session cache that syncs in the background. recall queries the graph and auto-routes to the best search strategy. forget(dataset=...) deletes a dataset. improve(...) runs the truth-subspace build after session distillation and before enrichment; v1.2.2 added the build_truth_subspace=True opt-in flag. The CLI mirrors the same surface: cognee-cli remember, cognee-cli recall, cognee-cli forget --all, and cognee-cli -ui to open the local graph UI.

Single Postgres is the recommended default, but optionality remains. Per the README’s “Run the Whole Memory Layer on Postgres” section, “In cognee 1.0 you can run the entire memory layer on a single Postgres instance” (README). The traditional agent-memory stack needs four services — a graph database, a vector database, Redis, and a relational database — all deployed before an agent remembers anything. cognee 1.0 collapses that to one Postgres instance with pgvector for embeddings, a SQL session-cache backend, and the same Postgres for metadata. The README reports that “in our CI benchmarks, Postgres search ran ~10% faster than the separate graph-plus-vector setup.” Optional dedicated backends remain available: Neo4j and Neptune for graphs, Redis for sessions, pgvector and LanceDB for vectors, plus Qdrant, ChromaDB, Weaviate, and Milvus via community adapters. Local development stays fully embedded — SQLite, LanceDB, Kuzudb — with no extra services. Install is pip install "cognee[postgres]" with DB_PROVIDER=postgres, VECTOR_DB_PROVIDER=pgvector, GRAPH_DATABASE_PROVIDER=postgres, CACHE_BACKEND=postgres.

The research grounding is a real, dated arXiv paper — but a preliminary version. Optimizing the Interface Between Knowledge Graphs and LLMs for Complex Reasoning (Markovic, Obradovic, Hajdu, Pavlovic; submitted 2025-05-30 11:27:59 UTC; subjects cs.AI, cs.CL; DOI 10.48550/arXiv.2505.24478) studies hyperparameter optimization in the context of Cognee across three multi-hop QA benchmarks — HotPotQA, TwoWikiMultiHop, and MuSiQue — scored with exact match, F1, and DeepEval’s LLM-based correctness metric. The paper’s own summary: “Our results demonstrate that meaningful gains can be achieved through targeted tuning. While the gains are consistent, they are not uniform, with performance varying across datasets and metrics.” The arXiv page is explicit: “This is a preliminary version. A revised and expanded version is in preparation” — the article cites the paper accordingly.

The BEAM benchmark is self-reported and explicitly directional. Per the README, the BEAM table reads cognee 0.79 at 100K tokens (>0.8 with per-question routing) vs previous SOTA 0.735, and 0.67 at 10M tokens vs previous SOTA 0.641, with an Obsidian / RAG baseline of ~0.33. The README’s own caveat, reproduced verbatim: “These numbers are a directional signal rather than a definitive measure — see the write-up for the full methodology, caveats, and what the results actually mean.” No third-party reproduction is on the record as of 2026-06-30.

The Claude Code plugin is the most direct builder hook. Cognee ships a Claude Code marketplace plugin — cognee-memory@cognee — that “captures prompts, tool traces, and assistant responses into session memory, injects relevant context on every prompt, and syncs session memory into the permanent knowledge graph at session end” (plugin source). The lifecycle hooks cover SessionStart, UserPromptSubmit, PostToolUse, Stop, PreCompact, and SessionEnd. Install is three commands before the first claude launch:

claude plugin marketplace add topoteretes/cognee-integrations
claude plugin install cognee-memory@cognee
export LLM_API_KEY="sk-..."   # local mode; or COGNEE_BASE_URL + COGNEE_API_KEY for cloud
claude

On startup, Claude Code shows a “Cognee Memory Connected” system message. For Cognee Cloud or a remote instance, the plugin points at https://your-instance.cognee.ai with a ck_... API key.

The MCP server attaches cognee to any agent that speaks MCP. Pre-built images are published as cognee/cognee (API server on port 8000) and cognee/cognee-mcp (MCP server on port 8001). Either docker compose --profile mcp up or docker pull cognee/cognee-mcp:main with TRANSPORT_MODE=http exposes cognee as a first-class MCP server. Any MCP-speaking agent — Claude Code, Codex, Cursor, OpenCode, Cline, Windsurf, Continue, Kiro — can attach a persistent memory layer without writing custom glue code.

Other clients and add-ons round out the language surface. Beyond Python, cognee ships official clients for Rust (cognee-rs, cargo add cognee) and TypeScript (@cognee/cognee-ts, npm install @cognee/cognee-ts); an OpenClaw plugin (@cognee/cognee-openclaw); and Cognee Cloud at www.cognee.ai. One-click deploys target Modal, Railway, Fly.io, Render, and Daytona.

Distinct category from existing AI Newsroom coverage. The closest existing AI Newsroom piece tagged knowledge-graph is codebase-memory-mcp (2026-06-24) — a code-intelligence MCP for code search, not a cross-session agent memory layer. cognee and codebase-memory-mcp solve different problems; both can coexist in the same agent stack.

v1.2.2 (2026-06-26) is the headline release: truth-subspace reranking, all opt-in. Per the v1.2.2 release notes, the headline feature is a truth subspace — a compact index built from distilled, accepted session learnings (centroids as topic anchors, slots as grouped examples) that helps rerank search results. The release ships five named items:

  1. Truth subspace builder — compiles distilled session lessons into centroids and slots.
  2. Centroid-slot truth weighting (MVP) — adjusts ranking scores using the truth subspace.
  3. Truth-subspace reranking + feedback activation — reorders search results to favor items aligned with prior accepted lessons; tunable per call.
  4. build_truth_subspace opt-in flag on /improve — runs the truth-subspace build after session distillation and before enrichment.
  5. DEFAULT_FEEDBACK_INFLUENCE env var — controls the default learned-feedback blend weight during graph search; default is 0.0 (i.e. reranking does not change behavior unless the operator opts in).

Two demos are new: examples/demos/truth_centroid_slots_demo.py and examples/python/truth_subspace_reranking_demo.py. New abstract methods get_node_truth_state and set_node_truth_state were added to the graph DB interface. Truth signatures now use sha256 hashing. The LanceDB S3 usage bug is fixed. Breaking changes: none — all new behavior is opt-in by default.

v1.2.0 (2026-06-21) is the security-and-distillation release with two breaking env var renames. Per the v1.2.0 release notes, the headline is smart session distillation — auto-distill during improve(), distilled nodes marked in the visualization, and a reworked distillation pipeline (batched curator, per-lesson writers, reuse of stored embeddings, one document per lesson). The breaking changes are LLM_MAX_TOKENS → LLM_MAX_COMPLETION_TOKENS and EMBEDDING_MAX_TOKENS → EMBEDDING_MAX_COMPLETION_TOKENS. Public registration is disabled by default; ENABLE_BACKEND_ACCESS_CONTROL now implies API auth and per-user/dataset DB isolation. v1.2.0 also adds a Proposals endpoint, inline skill ingest API, S3FileStorage boto3 credential chain fallback (IAM role support), LanceDB Windows long-path prefixing, and LLM/embedding resiliency (retry capping, over-length embedding handling, OpenAI fallback endpoints). v1.2.1, tagged the same day, ships “Reliability & Search Improvements” — small fixes to LLM request handling. Operators tracking the cadence should note that v1.2.0 and v1.2.1 landed within hours of each other.

Why it matters

1. Agent memory is the unsolved infrastructure problem of agent tooling. cognee is the leading open-source answer — a self-hosted, Apache-2.0, graph-backed memory layer that any agent can attach via SDK, CLI, MCP, or marketplace plugin.

2. The single-Postgres architecture is a real simplification, not marketing. The traditional agent-memory stack needs four deployed services before an agent remembers anything. cognee 1.0 collapses that to one Postgres instance, and the README reports the same Postgres setup ran ~10% faster than the separate graph-plus-vector setup in CI benchmarks. The trade-off is that the recommendation is now Postgres-first, not Postgres-only — the README’s architecture table still lists Neo4j, Neptune, Redis, pgvector, LanceDB, Qdrant, ChromaDB, Weaviate, and Milvus as available dedicated backends.

3. The arXiv paper is real research grounding, with honest limits. Markovic et al.’s paper studies hyperparameter optimization across three standard multi-hop QA benchmarks — HotPotQA, TwoWikiMultiHop, MuSiQue — scored with exact match, F1, and DeepEval’s LLM-based correctness metric. The paper is explicit about its own limits — “the limitations of standard evaluation measures” and “performance varying across datasets and metrics” — and the arXiv page flags the paper as a preliminary version with a “revised and expanded version in preparation.” The article preserves both halves of the framing.

4. The Claude Code plugin is the lowest-friction entry point on the market today. Most “agent memory” platforms require custom glue code. cognee ships a Claude Code marketplace plugin that captures prompts, tool traces, and assistant responses into session memory, injects dataset-scoped context on every prompt, and syncs session memory into the permanent knowledge graph at session end. The install path is three commands and the value is visible in the next session.

5. v1.2.2’s truth-subspace reranking is a research-grade retrieval improvement, shipped behind a default-off flag. The build_truth_subspace opt-in flag, the DEFAULT_FEEDBACK_INFLUENCE env var (default 0.0), the new graph-DB interface methods, the new demos, and the sha256 signature hashing are all in v1.2.2. The README does not claim “truth-subspace reranking improves accuracy by X%” — no such claim is on the record as of 2026-06-30. The default DEFAULT_FEEDBACK_INFLUENCE of 0.0 does not change retrieval behavior.

Practical implications for builders and operators

1. If you are running Claude Code today, install the cognee plugin in three commands. claude plugin marketplace add topoteretes/cognee-integrations, claude plugin install cognee-memory@cognee, then export LLM_API_KEY="sk-..." and launch Claude Code. The first launch bootstraps memory automatically.

2. If you are running any MCP-speaking agent, attach cognee-mcp via Docker. docker pull cognee/cognee-mcp:main && docker run -e TRANSPORT_MODE=http --env-file ./.env -p 8000:8000 --rm -it cognee/cognee-mcp:main exposes the cognee API as an MCP server on localhost:8000. Configure the agent’s MCP client to point at the container.

3. If you are deploying for the first time, default to single Postgres. pip install "cognee[postgres]", then set DB_PROVIDER=postgres, VECTOR_DB_PROVIDER=pgvector, GRAPH_DATABASE_PROVIDER=postgres, CACHE_BACKEND=postgres, plus the DB_HOST / DB_PORT / DB_USERNAME / DB_PASSWORD / DB_NAME vars. The README reports ~10% faster retrieval than the separate graph-plus-vector setup in CI benchmarks.

4. If you are upgrading to v1.2.0 or later, rename two env vars. LLM_MAX_TOKENS → LLM_MAX_COMPLETION_TOKENS and EMBEDDING_MAX_TOKENS → EMBEDDING_MAX_COMPLETION_TOKENS. Skipping this rename silently disables your completion-token cap. Review your .env and your deployment configs before rolling out v1.2.0+.

5. If you want to try truth-subspace reranking, opt in per-call or per-deployment. Set DEFAULT_FEEDBACK_INFLUENCE to a non-zero value (e.g. 0.1) to enable the learned feedback signal globally, or pass feedback_influence per API call. Then call improve(..., build_truth_subspace=True) to build the index after session distillation. Run the demos (truth_centroid_slots_demo.py, truth_subspace_reranking_demo.py) to verify behavior on your data.

6. If you are building an agent in Python, the SDK path is four lines of code. uv pip install cognee, then import cognee; await cognee.remember(...), await cognee.recall(...), await cognee.forget(...). For Rust, use cognee-rs (cargo add cognee); for TypeScript, use @cognee/cognee-ts (npm install @cognee/cognee-ts); for OpenClaw, install @cognee/cognee-openclaw.

Risks and caveats

  1. The BEAM numbers are self-reported and explicitly directional. Per the README, verbatim: “These numbers are a directional signal rather than a definitive measure — see the write-up for the full methodology, caveats, and what the results actually mean.” The article does not present cognee’s 0.79 / 0.67 as a definitive ranking. No third-party reproduction of the BEAM numbers is on the record as of 2026-06-30.

  2. The arXiv paper is a preliminary version, not a peer-reviewed publication. The arXiv page states, verbatim: “This is a preliminary version. A revised and expanded version is in preparation.” The paper is cited as arXiv:2505.24478, not as a peer-reviewed conference paper. The paper is honest about its own limits — “the limitations of standard evaluation measures” and “performance varying across datasets and metrics” — and the article preserves both halves of the framing.

  3. Truth-subspace reranking is opt-in and the new behavior. Per the v1.2.2 release notes, the default DEFAULT_FEEDBACK_INFLUENCE is 0.0 — i.e. reranking does not change behavior unless the operator sets DEFAULT_FEEDBACK_INFLUENCE or passes a non-zero feedback_influence per API call. The article does not characterise truth-subspace reranking as “now active by default” — it is now available, opt-in. Existing deployments that upgrade to v1.2.2 do not see a behavior change unless they opt in.

  4. Single Postgres is the recommended default, but optionality remains. The README’s architecture table lists Neo4j and Neptune for graphs, Redis for sessions, pgvector and LanceDB for vectors, plus Qdrant, ChromaDB, Weaviate, and Milvus via community adapters. The article does not claim cognee is “Postgres-only” — it is Postgres-first, with optional dedicated backends.

  5. v1.2.0 introduced a breaking env var rename. Operators relying on LLM_MAX_TOKENS or EMBEDDING_MAX_TOKENS must update their .env or deployment configs to the new LLM_MAX_COMPLETION_TOKENS / EMBEDDING_MAX_COMPLETION_TOKENS names. v1.2.0 also disabled public registration by default and tightened ENABLE_BACKEND_ACCESS_CONTROL to imply API auth and per-user/dataset DB isolation — operators who relied on the previous default-open posture must review their auth settings.

  6. No independent head-to-head against a comparable agent-memory platform. Zep, mem0, Letta, and other agent-memory platforms exist; the cognee README and the AI Newsroom radar sweep do not surface a head-to-head benchmark, and no such benchmark is on the record as of 2026-06-30. Cognee Cloud pricing is also not on the public record as of 2026-06-30 — operators evaluating cloud versus self-hosting should contact the vendor.

What to watch

  1. Revised arXiv paper. A revised and expanded version of arXiv:2505.24478 is in preparation per the arXiv page.
  2. First independent BEAM benchmark reproduction. No third-party reproduction is on the record as of 2026-06-30. The first academic or community-led reproduction of the 100K / 10M BEAM numbers will be a separate article.
  3. Cognee Cloud GA / pricing. Cognee Cloud exists (www.cognee.ai), but a published pricing page is not on the record as of 2026-06-30. The first published cloud pricing or self-serve GA announcement will be a separate article.
  4. v1.3 / v1.4 with truth-subspace as default. v1.2.2 ships truth-subspace as opt-in; the v1.3 / v1.4 line will likely turn it on by default. Watch the cognee releases page.
  5. Adoption in non-Claude-Code agents. The MCP server (cognee/cognee-mcp) is the hook for any MCP-speaking agent. The first major non-Claude-Code agent to ship a first-party cognee integration will be a follow-up article.

Sources