Persistent, claim-focused memory server for MCP-enabled AI agents
tensory, from Kryptogrib, is a memory server that gives AI agents long-term, verifiable context. It extracts atomic claims from conversations and documents and exposes structured memory to models. Key aspects include claim extraction, contradiction detection, and a searchable memory interface. The tool targets AI developers and engineers building stateful agents who need persistent, auditable context without heavy infrastructure.
What tasks can you actually use it for?
The server supplies persistent context for agents operating across sessions, so agents can reference past facts and detect inconsistencies. It integrates with MCP-compatible clients such as Claude Desktop, Cursor, and Zed, which lets coding assistants and chat agents query stored facts during prompts. Use cases include multi-session assistants, developer tooling that needs prior user preferences, and agents that must track evolving project facts over time.
How accurate are the memories and retrieval?
Accuracy is anchored to its claim-native approach and benchmark performance: tensory scored 82.2% on the LoCoMo long-term conversational memory benchmark. Retrieval combines full-text FTS5 search, vector embeddings, and graph traversal to match queries to stored statements. Built-in collision detection identifies contradictory or superseded claims, which helps keep retrieved context aligned with later information.
What inputs, requirements, and workflow fit should you expect?
The server runs where an MCP client can reach it and requires Python 3.11 or newer for deployment. It stores memory in a single SQLite-based graph and vector store, so it operates without external database services. The package includes a web dashboard for exploring entity graphs and memory statistics, and internal mechanisms such as salience decay, surprise scoring, and priming operate without extra calls to the language model.
tensory suits developers who need verifiable, long-lived agent memory
The server is a practical option for teams building stateful agents that require auditable facts and contradiction handling, and it is recognized within the MCP developer community for its precision. Expect integration work to connect MCP clients and adapt claim extraction to your domain, and treat stored claims as a source to verify for high-stakes decisions.
