Cogni
MCP memory for LLMs, with no LLM in the retrieval path.
Last verified:
What is Cogni?
Cogni is an MCP server that provides LLMs with connected, reasoning memory using entity-graph spreading activation. It retrieves information across documents even when queries and answers don't share vocabulary—solving multi-hop reasoning problems that vector stores can't reach in a single call. Works with any LLM model without requiring a vector database, GPU, or LLM in the retrieval path.
Cogni pricing
Pricing model: Freemium
Free tier available ("Get started free"); detailed pricing not specified on page
Cogni pros
- Excels at cross-vocabulary and multi-hop retrieval (0.85 recall vs 0.65 for vector search on chain-recall benchmarks)
- Works with any LLM (Claude, GPT, Gemini, local) with no LLM or GPU required for retrieval
- Deterministic, reproducible results—same question returns same rows every time, no model variability
- No infrastructure to maintain: no vector database, GPU, or external API keys needed; includes full vector search as a floor
- Recall performance stays flat as memory grows to 32k+ words—effectively a rolling context window at constant per-query cost
Cogni cons
- No advantage on simple single-document retrieval (matches vector search at 1.00 to 1.00 accuracy)
- Still requires the calling LLM to correctly invoke tools and be prompted appropriately for multi-hop reasoning
- Performance depends on chain length—longer chains show more benefit, but still require deterministic traversal
Frequently asked questions about Cogni
How does Cogni differ from vector search?
Cogni uses entity-graph spreading activation to traverse chains of facts across documents. Vector search uses similarity matching and scores zero on cross-vocabulary queries where the answer shares no words with the question. On simple single-document retrieval, both perform equally.
Do I need a vector database or GPU?
No. Cogni requires no vector database, GPU, or additional LLM in the retrieval path. It performs deterministic graph traversal and includes a full vector search layer underneath.
What's the cost advantage vs. vector search?
Vector search reaches multi-hop answers through iterative calls (each an API round trip and context window charge). Cogni returns the chain in a single call, reducing total API costs as chain length increases.
Does it work with my LLM?
Yes—Cogni works with any model including Claude, GPT, Gemini, and local models. The retrieval advantage is deterministic and doesn't depend on LLM choice.