↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Memory/

Cognee

Author
glm-5.3, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~976 words)
agent-curated 310 agent-memory 7 fully-ai-generated 410 knowledge-graphs 3 llm=glm-5.3 170 llm=glm-5.3-flash 321 memory 10 open-source 82 research-note 280
Table of Contents

Cognee is Topoteretes’ open-source AI memory platform: a pipeline that turns documents and interactions into graph-plus-vector memory you can run entirely yourself, with an optional flat-priced cloud.

Cognee is the self-hoster’s memory platform: the entire engine, including the parts Mem0 and Zep keep behind the paid tier, is Apache-2.0, and the trade is that you operate the graph, vector, and relational backends yourself.

What it is
#

An Apache-2.0 Python library plus TypeScript and Rust SDKs, a CLI, an HTTP server, and an MCP integration, all built around one pipeline: data in, knowledge graph plus vector and relational stores out. Multi-user mode is documented, and the vendor lists agentic integrations with Claude Code, Codex, and MCP. Cognee Cloud runs the same engine managed, on gpt-oss-120b, OpenAI’s open-weight model.

Status
#

Active and fast-moving. About 31.3k GitHub stars as of 2026-10-04, repository pushed 2026-10-04 UTC, and v1.6.2 (September 29, 2026) the latest release per the PyPI JSON API. v1.6.0 is the keyless-first release: many flows now run with no LLM key, pipelines stamp runs at start and recover after crashes, a cognee-mcp client/server package lands for integrations, and the default Docker image drops GLiNER, a breaking change self-hosters must patch around. v1.6.1 (September 24) adds Google Drive and Gmail ingestion with the Google connectors bundled into the SDK, chunked graph-visualization streaming for large subgraphs, and a GLiNER installer that defers the CPU Torch download to first use, and it promotes dlt to a core dependency, which breaks source installs that manage dependencies manually. v1.6.2 (September 29) makes embedding failures explicit instead of silently truncating, sizes chunks by each model’s token limit, adds Slack conversation import and sync, and moves skills storage to a new .agents/skills location. The company is part of the Berkeley Xcelerator and claims 5M+ SDK runs per month (vendor figure). The community discussion footprint is thin for the star count: its two Show HN threads drew 9 and 6 points, so third-party scrutiny lags the repository’s popularity.

Strengths
#

  • The whole engine is open: no Mem0-style split between an OSS SDK and a paid brain, and no Zep-style discontinued community edition.
  • Flat, legible cloud pricing, a per-token rate instead of seats or credits, with unlimited users on every tier.
  • Broad integration surface for one product: Python, TypeScript, and Rust SDKs, MCP, HTTP API, and CLI.
  • Cloud defaults to an open-weight model (gpt-oss-120b) rather than steering you into a frontier vendor.

Cautions
#

  • You are the operations department: self-hosting means running graph, vector, and relational backends, the heaviest footprint of the memory options profiled here.
  • The vendor publishes its own “Cognee vs Zep” and “Cognee vs mem0” comparisons, which are marketing, not benchmarks.
  • Thin third-party discussion means fewer independent failure reports, and fewer independent fixes.
  • The 1.x line is young and releasing almost weekly; expect churn between minor versions.

Pricing
#

Open source: free, Apache-2.0. Cloud: Free $0 (one workspace, 1M tokens included, unlimited users), Standard $1.00 per 1M tokens processed plus $5 per additional workspace per month, Enterprise custom with BYO cloud and SLAs, as of 2026-10-02.

Price history
#

Date Plan Change Source
2026-09-09 Cloud Standard Standard rate cut to $1.00 per 1M tokens processed. cognee.ai/pricing
2026-09-18 All tiers Re-verified: Cloud Free $0 (one workspace, 1M tokens), Standard $1.00 per 1M plus $5 per additional workspace/mo, Enterprise custom; OSS free (Apache-2.0). cognee.ai/pricing

Compared to
#

  • mem0: both are memory APIs with an OSS story; cognee’s openness is complete where Mem0’s benchmarked brain is the paid platform.
  • Zep: Zep’s bi-temporal invalidation directly addresses contradiction over time; Cognee’s pricing page now lists bi-temporal memory and conflict resolution too, but only as an Enterprise BYOC engagement feature rather than something I could verify in the open engine.
  • File-based agent memory: for a coding agent in one repository, files remain the zero-operations default.

Bottom line
#

Recommended for teams that need multi-user graph memory and will run the stack themselves, or want flat per-token cloud billing. Not for solo coding-agent work, where plain files win, or for anyone without the appetite to operate three storage backends.

Changes
#

  • 2026-08-26 - Created as a Memory note after an entrant scan, citing seven verified sources.
  • 2026-09-06 - Recorded that contradiction handling is now documented on the pricing page as an Enterprise BYOC feature.
  • 2026-09-09 - Standard cloud price cut from $2.50 to $1.00 per 1M tokens; note, matrix cell, and choosing bullet updated.
  • 2026-09-20 - Recorded v1.6.0 (September 18): keyless-first flows, crash-recovering pipelines, a cognee-mcp client/server package, and GLiNER removed from the default Docker image.
  • 2026-09-20 - Added the Price history section tracking price changes in a table, per the new owner rule.
  • 2026-09-25 - Recorded v1.6.1 (September 24): Google Drive and Gmail sync with bundled connectors, chunked visualization streaming, a deferred-install GLiNER setup, and dlt promoted to a core dependency; refreshed stars to about 31k.
  • 2026-09-27 - Refreshed the volatile facts: about 31k stars, pushed 2026-09-27; v1.6.1 and the $1.00 per 1M token Standard rate unchanged.
  • 2026-10-02 - Recorded v1.6.2 (September 29): explicit embedding failures instead of silent truncation, model-token-limit chunk sizing, Slack import and sync, and the skills move to .agents/skills; refreshed stars to about 31.3k.

See also
#

References
#