mentedb

MenteDB is an open‑source, Rust‑based database that acts as the long‑term memory layer for AI agents. It is released under the Apache 2.0 license and can be used either through a hosted endpoint (you obtain an mdb_ API key) or by running the engine yourself after installing the mentedb package with pip or launching the provided server binary.

The core workflow revolves around a single call per interaction, named process_turn. You send the latest user or system utterance to this endpoint; the engine parses the text, extracts atomic facts such as statements of preference, decisions, or corrections, and stores them without requiring you to define a schema. New facts are automatically deduplicated, stale items decay over time, and any contradictions are resolved by linking the items into a knowledge graph. When the agent needs context for the next response, MenteDB performs a one‑pass hybrid retrieval: it selects a bounded set of the most relevant memories, reranks them, and returns the assembled context as a plain text block that can be appended to the model's prompt. This retrieval adds no extra per‑query token tax, keeping the cost per turn flat even as the underlying memory grows.

Semantic search is available across sessions. You can query the memory with natural language, e.g., “What did we decide about authentication?”, and the system uses embeddings to match the intent to stored facts, returning the matching snippets rather than relying on exact string matching. Contradiction detection runs continuously; if the agent previously asserted “Postgres” as the database and later says “MySQL,” MenteDB flags the inconsistency so you can intervene. Pain‑warning alerts surface when the current context contains patterns that previously led to errors, helping prevent repeated mistakes. Cloud sync keeps the same memory graph available on multiple devices, so a laptop and a desktop see identical memories without manual merging.

The product is positioned for developers building autonomous or semi‑autonomous agents that need persistent, searchable memory beyond a single context window. It suits use cases where agents must remember policies, user preferences, or evolving configurations over many turns. It is less appropriate for simple chatbots that operate entirely within a single request‑response cycle, because the added storage and retrieval steps provide little benefit in that scenario.