engrava-mcp

Engrava is a memory database designed specifically for AI agents, providing a local, embedded solution for managing agent memory with features such as graph memory, hybrid search, and an optional tamper-evident journal. It operates on SQLite and does not require a separate server, allowing all operations to run within the user's Python process. The platform is intended for developers building AI agents who need a deterministic, typed, and local data layer for storing and retrieving structured memory.

The database supports hybrid search by combining FTS5 (full-text search), vector search, recency, priority, and graph-based queries, fusing five different signals for retrieval. Engrava includes primitives for storing and recalling memories with simple function calls, and it features auto-embedding of data without requiring manual vector management or record assembly. The system manages memory structure using typed thoughts and typed edges, all contained within a single SQLite file. It also provides lifecycle management for data, with states such as CREATED, ACTIVE, DONE, and ARCHIVED.

Additional capabilities include an algorithmic memory consolidation feature called "dreaming," which operates deterministically based on configurable signals and gates, without relying on language model calls. The optional audit trail records mutations in a SHA-256 hash-linked journal, making the journal tamper-evident. Engrava offers a query language called MindQL, supporting commands like FIND, COUNT, and SELECT, and allows bi-temporal queries using valid-time attributes for facts, enabling time-travel queries and logical invalidation without deletion. Multiple embedding providers are supported, including local models, OpenAI-compatible services, Ollama, and HuggingFace.

Engrava can be installed via pip and configured with a YAML file. It includes a standalone MCP server (engrava-mcp) for connecting agents and offers a read-only mode. The tool supports isolated memory per service, with each managed in its own database file. Engrava is distributed under the MIT license, and there is no required service, per-operation metering, or data egress unless a remote provider is configured.