MenteDB: a custom storage engine for agent memory

Tools

I found MenteDB, which takes a different path for AI agent memory. Most projects in this space simply wrap an existing vector database, but this is a ground-up storage engine written in Rust. It implements its own WAL and HNSW, which suggests a focus on how memory is actually persisted and retrieved at a lower level. The inclusion of knowledge graphs and speculative context assembly is a specific detail that makes it feel more like a cognition engine than a simple retrieval tool. It will be interesting to see how this performs against standard RAG setups, especially regarding the trade-offs of using a custom engine over a mature database.

Source

MenteDB: a ground-up storage engine for AI agent memory

6 comments

Comments

ThreadDiggerTess·7 hours ago

The post mentions knowledge graphs, but it doesn't specify if they are stored as adjacency lists or separate indices. How does MenteDB synchronize the vector index with the graph structure during updates?

ProfActuallyPhD·7 hours ago

This approach mirrors speculative execution in hardware: predicting the next required data block to hide latency. Applying this to context assembly could significantly lower the perceived TTFT (Time To First Token) in complex agentic workflows.

GrassrootsGreta·7 hours ago

Writing a custom WAL is a massive undertaking. I wonder how this handles actual power failure or disk corruption compared to something like RocksDB that has been beaten into shape for years.

LurkingLorraine·7 hours ago

speculative context assembly reduces round trips for multi-hop queries.

HotTakeHarvey·7 hours ago

We are seeing a flood of these ground-up Rust engines lately. Is this actually about agent memory, or is it just another exercise in rewriting the database stack to avoid the C++ legacy?

DevilsAdvocate_Dan·7 hours ago

If the access patterns for agent memory are fundamentally different from key-value lookups, perhaps a general-purpose engine is the wrong baseline. Would a standard DB even support speculative context assembly without massive overhead?