Postgres + Open-Source Models Take On Company Ontology as Cognee Hits 91% Accuracy
Cognee is an open-source tool that connects data sources such as GitHub, Slack, and Notion to automatically build a company ontology for AI agents. It does not use a graph database and does not rely on Anthropic/OpenAI; it runs only on Postgres and open-weight models. It achieves 91% accuracy on LoCoMo and 79% on BEAM.
AI agents now have to handle an ever-growing volume of data, but they lack an understanding of the relationships between pieces of information. To answer a simple question, they may have to go through the entire knowledge base and burn through millions of tokens.
Cognee aims to change that. It is an open-source tool that can connect more than 10 data sources—including GitHub, Slack, Linear, and Notion—to your agents or software factory, automatically building a company's ontology, in other words, structuring and organizing the relationships between pieces of information.
Unlike common approaches, Cognee does not use a graph database and does not depend on proprietary models from Anthropic or OpenAI. It uses small open-weight models to build different memory elements, then uses modified embeddings to avoid a graph database, with everything running on Postgres. In other words, an existing company may be able to have an ontology layer without adding a complex new infrastructure stack.
In terms of results, Cognee reaches 91% accuracy on LoCoMo and 79% on BEAM. These two benchmarks mainly test long-term memory and relationship understanding. The numbers are not top-tier, but considering that it uses only Postgres and open-source models, the result is compelling.
The video demonstrates the specific workflow:
The core problem it solves is cost. When an agent needs to understand relationships between pieces of information, rather than having it repeatedly read files, it is better to pre-build a lightweight ontology. Cognee's value lies in automating this process without locking you into a single model vendor.
If you are building a memory layer or knowledge layer for AI agents, this project is worth a look. Free, Postgres, open source—those words together are at least worth a quick run.
发布时间: 2026-11-20 03:33