cognee is the agent memory stack hitting 15K stars
Agent memory has been a mess for two years. Everyone rolls their own — vector DB here, Postgres there, some kind of graph walk. Cognee is the first serious attempt at a standard. Open source, sits at 15.8K GitHub stars, and you can get persistent agent memory running in six lines of code.
The design is Remember, Recall, Forget, Improve. Remember persists what the agent learns. Recall auto-routes between vector search and graph traversal depending on the query. Forget deletes — the missing primitive in most agent memory systems, because nobody wants to think about deletion until they need GDPR compliance. Improve updates the knowledge structure as new data comes in. Four verbs, one clean mental model.
Under the hood it's vector search and graph DB fused together, with session memory for fast caching plus permanent graph storage for long-term patterns. Multimodal, runs locally or in the cloud. Already wired into Claude Code and Hermes Agent.
The use case that sells this is boring: customer support agents that actually remember the customer between calls. But the deeper bet is that agents without memory don't compound. A memoryless agent does the same thing well 1000 times. An agent with real memory gets better the 1001st time. That's where the leverage is. https://github.com/topoteretes/cognee
← Back to all articles
The design is Remember, Recall, Forget, Improve. Remember persists what the agent learns. Recall auto-routes between vector search and graph traversal depending on the query. Forget deletes — the missing primitive in most agent memory systems, because nobody wants to think about deletion until they need GDPR compliance. Improve updates the knowledge structure as new data comes in. Four verbs, one clean mental model.
Under the hood it's vector search and graph DB fused together, with session memory for fast caching plus permanent graph storage for long-term patterns. Multimodal, runs locally or in the cloud. Already wired into Claude Code and Hermes Agent.
The use case that sells this is boring: customer support agents that actually remember the customer between calls. But the deeper bet is that agents without memory don't compound. A memoryless agent does the same thing well 1000 times. An agent with real memory gets better the 1001st time. That's where the leverage is. https://github.com/topoteretes/cognee
Comments