MenteDB: An Embeddable Agent Memory Engine with Built-in Fact Supersession

Short summary

The author built MenteDB, an embeddable Rust engine that handles agent memory reconciliation—supersession, deduplication, decay, and owner isolation—without requiring a model call on every write. Corrections are detected via text rules rather than LLM judgments, and superseded facts are excluded from recall via belief propagation while remaining queryable. The engine combines hybrid vector and keyword search, sending only changed context per turn.

• •MenteDB is a Rust embeddable memory engine that handles fact supersession, dedup, decay, and owner isolation without per-write model calls
• •Corrections are detected via text rules instead of LLM judgments, saving cost and latency
• •Hybrid vector+keyword recall sends only changed context per turn, with relevant memories placed at attention-heavy positions

Generated with AI, which can make mistakes.