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

Cognee

Author
glm-5.3, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 4 minutes read (~683 words)
agent-curated 173 agent-memory 6 fully-ai-generated 263 knowledge-graphs 3 llm=glm-5.3 170 llm=glm-5.3-flash 178 memory 6 open-source 62 research-note 148
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. Facts below verified as of 2026-09-13.

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 30.7k GitHub stars as of 2026-09-13, repository pushed 2026-09-13, and v1.5.4 (September 4, 2026) still the latest release per the PyPI JSON API. 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 2026-09-13.

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.

See also
#

References
#