Cognee 1.0 Collapses the Entire Agent Memory Stack Onto One Postgres
Cognee hit 1.0 this week and the surge is real, hundreds of stars a day. It's an open-source memory platform that gives agents persistent long-term memory across sessions by quietly building a knowledge graph as they work. The API is built around four verbs that read like how memory actually behaves: remember, recall, improve, forget. Roughly 6 million memories get created on it every month across 100-plus companies, and Bayer runs it inside scientific research workflows.
The 1.0 headline is an architecture flex that anyone who's tried to build an agent memory layer will feel in their bones. Until now the standard recipe was a graph database for relationships, plus a vector store for semantic search, plus Redis for sessions, three separate systems to deploy, sync and pay for. Cognee 1.0 runs the whole thing, graph, vectors, sessions and metadata, on a single Postgres instance. The pile of infrastructure just collapsed into the boring database you already have.
Under the hood it splits memory into two layers that mirror how brains work: session memory as fast short-term working context, and permanent memory as long-term artifacts that keep getting cross-linked inside the graph as your knowledge evolves. It mixes vector embeddings, graph reasoning and cognitive-science-grounded ontologies, so your documents are both searchable by meaning and connected by relationships, and those relationships change over time instead of going stale.
We've covered a parade of agent memory products this quarter, Supermemory, Walrus, MemPalace, the Universal Memory Protocol, and the honest read is that memory is now a crowded category. So why cognee 1.0? Because the news isn't another memory store, it's the consolidation. The interesting move in a crowded category is the one that removes parts rather than adds them. Run-it-on-one-Postgres is a different pitch from rent-a-memory-cloud, and it's the kind of simplification that actually wins self-hosters.
Link: https://github.com/topoteretes/cognee
← Back to all articles
The 1.0 headline is an architecture flex that anyone who's tried to build an agent memory layer will feel in their bones. Until now the standard recipe was a graph database for relationships, plus a vector store for semantic search, plus Redis for sessions, three separate systems to deploy, sync and pay for. Cognee 1.0 runs the whole thing, graph, vectors, sessions and metadata, on a single Postgres instance. The pile of infrastructure just collapsed into the boring database you already have.
Under the hood it splits memory into two layers that mirror how brains work: session memory as fast short-term working context, and permanent memory as long-term artifacts that keep getting cross-linked inside the graph as your knowledge evolves. It mixes vector embeddings, graph reasoning and cognitive-science-grounded ontologies, so your documents are both searchable by meaning and connected by relationships, and those relationships change over time instead of going stale.
We've covered a parade of agent memory products this quarter, Supermemory, Walrus, MemPalace, the Universal Memory Protocol, and the honest read is that memory is now a crowded category. So why cognee 1.0? Because the news isn't another memory store, it's the consolidation. The interesting move in a crowded category is the one that removes parts rather than adds them. Run-it-on-one-Postgres is a different pitch from rent-a-memory-cloud, and it's the kind of simplification that actually wins self-hosters.
Link: https://github.com/topoteretes/cognee
Comments