openmem-engine 0.4.0

Cognitive memory engine for AI agents — human-inspired retrieval via activation, competition, and reconstruction

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.

How it works

Query → FTS5/BM25 (lexical trigger)
      → Seed Activation
      → Spreading Activation (graph edges, max 2 hops)
      → Recency + Strength + Confidence weighting
      → Competition (score-based ranking)
      → Context Pack (token-budgeted output)

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

Install

pip install openmem-engine

Or from source:

git clone https://github.com/yourorg/openmem.git
cd openmem
pip install -e ".[dev]"

Quick start

from openmem import MemoryEngine

engine = MemoryEngine()  # in-memory, or MemoryEngine("memories.db") for persistence

# Store memories
m1 = engine.add("We chose SQLite over Postgres for simplicity", type="decision", entities=["SQLite", "Postgres"])
m2 = engine.add("Postgres has better concurrent write support", type="fact", entities=["Postgres"])

# Link related memories
engine.link(m1.id, m2.id, "supports")

# Recall
results = engine.recall("Why did we pick SQLite?")
for r in results:
    print(f"{r.score:.3f} | {r.memory.text}")
# 0.800 | We chose SQLite over Postgres for simplicity
# 0.500 | Postgres has better concurrent write support

Claude Code plugin

One command to add persistent memory to Claude Code:

uvx openmem-engine install

That's it. Claude now has 7 memory tools (memory_store, memory_recall, memory_link, memory_reinforce, memory_supersede, memory_contradict, memory_stats) it can call automatically across sessions.

Memories persist in ~/.openmem/memories.db by default (override with the OPENMEM_DB env var).

Usage with an LLM agent

engine = MemoryEngine("project.db")

# Agent stores what it learns
engine.add("User prefers TypeScript over JavaScript", type="preference", entities=["TypeScript", "JavaScript"])
engine.add("Auth system uses JWT with 24h expiry", type="decision", entities=["JWT", "auth"])
engine.add("The /api/users endpoint returns 500 on empty payload", type="incident", entities=["/api/users"])

# Before each LLM call, recall relevant context
results = engine.recall("set up authentication", top_k=5, token_budget=2000)
context = "\n".join(r.memory.text for r in results)

prompt = f"""Relevant context from previous work:
{context}

User request: {user_message}"""

API

MemoryEngine(db_path=":memory:", **config)

Method	Description
add(text, type="fact", entities=None, confidence=1.0, gist=None)	Store a memory
link(source_id, target_id, rel_type, weight=0.5)	Create an edge between memories
recall(query, top_k=5, token_budget=2000)	Retrieve relevant memories
reinforce(memory_id)	Boost a memory's strength
supersede(old_id, new_id)	Mark a memory as outdated
contradict(id_a, id_b)	Flag two memories as contradicting
decay_all()	Run decay pass over all memories
stats()	Get summary statistics

Memory types

fact · decision · preference · incident · plan · constraint

Edge types

mentions · supports · contradicts · depends_on · same_as

Retrieval model

Recency — Exponential decay with ~14-day half-life. Recently accessed memories surface first.

Strength — Reinforced on access, decays naturally over time. Frequently recalled memories persist.

Spreading activation — Memories linked by edges activate their neighbors. A query hitting one memory pulls in related context up to 2 hops away.

Competition — Final score combines activation (50%), recency (20%), strength (20%), and confidence (10%). Superseded memories are penalized 50%, contradicted ones 70%.

Conflict resolution — When two contradicting memories both activate, the weaker one (by strength × confidence × recency) gets demoted.

Tests

pip install -e ".[dev]"
pytest tests/ -v

License

MIT