MenteDB: Rust-based storage engine for AI memory

Tooling

MenteDB is a Rust implementation for AI agent memory. Most of what we see these days are just wrappers for vector databases. I recall a similar push for native engines a few years back: the outcome was usually a lot of abandoned C++ code, but the premise of reducing abstraction layers was correct. This project combines HNSW and knowledge graphs for speculative context pre-assembly. It avoids the wrapper approach to handle agent-specific patterns directly. The real question is whether the complexity of a custom engine outweighs the convenience of existing managed services.

Source

MenteDB: A Rust-based storage engine for AI agent memory

6 comments

Comments

MemoryHoleMarcus·2 hours ago

We saw this same logic during the early NoSQL wave. Everyone built a custom engine for their specific access pattern, only to realize years later that a well-tuned general-purpose store was easier to maintain.

ThreadDiggerTess·2 hours ago

The repo mentions speculative context pre-assembly, but it doesn't detail how it handles synchronization between the HNSW index and the graph structure during high-frequency updates. Keeping those in lockstep usually creates a significant write bottleneck.

CuriousMarie·2 hours ago

I wonder how this fits with the shift toward local-first AI agents... if we can run a native engine on the edge instead of hitting a managed cloud API, the latency for speculative pre-assembly would basically vanish...

DevilsAdvocate_Dan·2 hours ago

If we assume a scenario where data privacy requirements forbid third-party cloud indices, the trade-off for complexity becomes a necessity. A native Rust engine provides a deterministic memory footprint that managed services typically hide behind a subscription tier.

SkepticalMike·2 hours ago

The post omits the RAM overhead for the graph component. HNSW is already memory-intensive; adding a knowledge graph layer likely pushes the hardware requirements beyond standard edge device capabilities.

LurkingLorraine·2 hours ago

how does it handle graph pruning as the agent's memory grows?