OpenMem

Deterministic memory engine for AI agents. Retrieves context via BM25 lexical search, graph-based spreading activation, and human-inspired competition scoring. SQLite-backed, zero dependencies.

No vectors, no embeddings, no LLM in the retrieval loop. The LLM is the consumer, not the retriever.

Why OpenMem?

Deterministic	Same input, same output. No stochastic embedding drift.
Zero dependencies	Pure Python + SQLite. Nothing to install, nothing to break.
Human-inspired	Memories decay, get reinforced, compete, and conflict — just like human recall.
Token-budgeted	Retrieval respects a token budget so you never blow your context window.

Install

pip install openmem-engine

Or add to Claude Code for persistent memory across coding sessions:

uvx openmem-engine install

Browse your memories in the Web UI:

openmem-engine ui

Quick example

from openmem import MemoryEngine

engine = MemoryEngine()

m1 = engine.add("We chose SQLite for simplicity", type="decision", entities=["SQLite"])
m2 = engine.add("Postgres has better concurrent writes", type="fact", entities=["Postgres"])
engine.link(m1.id, m2.id, "supports")

results = engine.recall("Why did we pick SQLite?")
for r in results:
    print(f"{r.score:.3f} | {r.memory.text}")