Cogni
Cogni is an MCP memory server that aims to provide LLM agents with a connected memory layer. It stores information as an entity graph and uses spreading activation during recall to follow chains of facts across documents, rather than relying only on lexical or vector similarity. The homepage describes it as a cognition layer for LLMs, available as a free hosted service and connectable to MCP-compatible clients such as Claude, ChatGPT, and other AI agents. The retrieval path is presented as deterministic graph traversal with no LLM and no GPU involved. Cogni's default recall combines a dense-vector search with the entity-graph spreading activation and returns both, so it can retain conventional vector retrieval while adding graph-based chain recall. The main claimed advantage is cross-vocabulary multi-hop retrieval, where the answer shares no words with the question and a single similarity lookup cannot reach it. Cogni is model-agnostic and works with hosted or local models, including small local models. It is offered with a free tier through a hosted MCP endpoint. The project reports benchmark improvements on multi-hop chain recall and local-model reasoning, but these are vendor-published results, and real-world performance depends on the calling model, prompting, and data.
- Category
- Model Serving
- Pricing
- Free
Problem solved
LLM agents and retrieval-augmented systems often rely on vector similarity search, which can fail on multi-hop or cross-vocabulary questions where the answer shares no lexical overlap with the query. Cogni addresses this by providing an MCP memory server with entity-graph spreading activation recall, connecting facts across documents so an agent can retrieve a reasoning chain in fewer round trips.
Use cases
- Add persistent, graph-based memory to Claude, ChatGPT, or other MCP-compatible LLM agents
- Answer cross-vocabulary multi-hop questions over a document store
- Reduce retrieval round trips in agentic RAG workflows
- Run memory and retrieval for local or small LLMs without a GPU in the retrieval path
- Combine dense vector search with entity-graph spreading activation for recall
- Build AI agents that need to connect facts across separate documents
Audience
- Developers
- ML Engineers
- Researchers
- Startups
Pros
- Works with any model and integrates via MCP
- No LLM or GPU in the retrieval path
- No separate vector DB required; vector search remains available underneath the graph
- Free hosted tier
- Deterministic graph traversal recall
Cons
- Newer tool with limited independent third-party validation
- Graph memory may require careful entity extraction and data modeling
- Advantages over iterative vector RAG narrow when agents can make multiple retrieval calls
- Benchmark results are vendor-reported from Cogni's own site
- MCP ecosystem compatibility depends on client support
Official website: https://getcogni.io/
Canonical LambdaBase page: https://www.lambdabase.com/tools/cogni