Most memory servers store whatever they are handed. This one scores every candidate first, and the scoring is pattern matching and statistics with no model call, so it costs nothing per write. The other engineering worth knowing is defensive: a circuit breaker sits in front of the FAISS and embedding path, and search degrades to BM25 rather than failing when embeddings are unavailable.
A Python memory server that writes to local JSONL with an optional FAISS index. Search fuses keyword and semantic scoring; every candidate memory is first scored on four dimensions by a purely algorithmic filter, with no model call involved.
- A memory persisted with `importance` from 1 to 4, tags and a confidence value, scored on the way in — `memory_store`
- Hybrid search combining BM25 at 40%, FAISS at 50%, importance at 5% and recency at 5%, falling back to keyword-only when embeddings are unavailable — `memory_recall`
- A context window assembled to a token budget — `memory_context`
- Confidence adjusted, a memory removed, and system statistics — `memory_update_confidence`, `memory_delete`, `memory_stats`
- The boundary score computed on demand, with acceptance rate and trigger statistics — `bnd_check`, `bnd_stats`
- Pattern extraction across high-quality memories — `deduce`
- Exact-match deduplication: an identical memory updates importance and confidence instead of storing a duplicate
No account and no key. `pip install mindcore-memory`, then run `mindcore-memory`, or launch the `mindcore_memory.server` module under `python` with `MINDCORE_MEMORY_PATH` pointing at your store. Semantic search is an optional extra — `pip install mindcore-memory[semantic]` — and without it search runs BM25-only; the `[encrypt]` extra adds Fernet encryption at rest. No PostgreSQL, Redis or external service is needed. The package is `mindcore-memory` 0.1.15.
One command — pip install mindcore-memory