MemX is a local-first long-term memory system for AI assistants. All data lives in a single file on your machine. Zero cloud. Zero tracking. Built with Rust.
// features
No subscription. No sync server. No account. Just a file on your disk.
Everything lives in one libSQL file. Backup by copying. Migrate by pasting. No config, no daemon, no database server.
Low-confidence queries return empty results instead of hallucinated answers. The rejection gate suppresses spurious recall at τ = 0.50.
Vector search + keyword search, fused via Reciprocal Rank Fusion. Gets the best of both semantic understanding and exact matching.
Four factors: semantic similarity, recency (half-life decay), retrieval frequency, and explicit importance. Memories stay relevant over time.
Every result is traceable. Retrieval stats are written back to the database so you can inspect exactly why a memory was returned.
Works with any OpenAI-compatible API — Ollama, LM Studio, or hosted providers. Swap models without changing your data.
// benchmark
Evaluated on two Chinese-language benchmark suites, 43 queries total, using real embeddings.
// how it works
A deterministic pipeline with no magic — every step is inspectable.
Query is sent to your local embedding model to produce a dense vector.
Vector search (cosine) and keyword search run in parallel against the memory store.
Reciprocal Rank Fusion merges both candidate lists into a single ranked set.
Four-factor scoring: semantic · recency · frequency · importance, then z-score normalized.
If both signals are weak, return ∅. Never return a bad answer.