openmem

How we compare

Other memory systems rely on embeddings and hosted infrastructure.
We think agent memory should be deterministic, local, and inspectable.

> openmemBM25 + graph

Mem0Vector, managed cloud

ZepKG + vector

LettaAgent runtime

File-basedCLAUDE.md, etc.

Deterministic retrieval

—

Graph relationships

—

Memory decay & reinforcement

—

Conflict resolution

—

Token-budgeted output

—

No embedding model needed

—

Local-first, no cloud

Inspectable scoring

—

N/A

Zero external dependencies

—

Open source

Six-stage retrieval

Every recall() runs a deterministic pipeline. No magic. No black boxes.

input

recall("how does auth work?")

01BM25FTS5 lexical match against text, gist, entities

02SeedNormalize BM25 scores to [0, 1]

03SpreadTraverse graph edges up to 2 hops with decay

04CompeteWeighted: activation + recency + strength + confidence

05ResolveDetect contradictions, demote weaker memory

06PackSort by score, fit within token budget

output

0.92 "Auth uses refresh tokens"0.87 "JWT tokens expire after 24h"0.41 "Session store uses Redis"

Built for agents

Everything your agent needs to remember, recall, and reason over time.

Deterministic

Same query, same results. No stochastic embedding drift. Every score is inspectable.

r.components
# {activation: 0.72, recency: 0.95,
#  strength: 0.80, confidence: 0.90}

Zero dependencies

Pure Python + SQLite. No vector DB, no embedding model, no external services.

engine = MemoryEngine()       # in-memory
engine = MemoryEngine("p.db") # persistent

Biological recall

Memories decay, get reinforced, compete for activation, and resolve conflicts.

engine.reinforce(m.id)     # boost
engine.supersede(old, new)  # outdated
engine.decay_all()          # forget

Token-budgeted

Retrieval respects a token budget. Never blow your context window.

engine.recall(
  query, top_k=10, token_budget=2000
) # packs by score until budget hit

Graph-linked

Spreading activation pulls in connected context that BM25 alone would miss.

engine.link(
  m1.id, m2.id, "supports", weight=0.8
) # m2 activates when m1 is recalled

Web UI

Browse, search, and inspect your memory store. One command.

openmem-engine ui
# Opens http://localhost:3333

MCP Integration

Drop into Claude Code

One command installs 7 MCP tools that give Claude persistent memory across every session. Your agent remembers what it learned yesterday.

storesave a memory

recallsearch & retrieve

linkconnect memories

reinforceboost strength

supersedereplace outdated

contradictflag conflicts

statsstore overview

Read the docs →

terminal
$ uvx openmem-engine install

Adding OpenMem to Claude Code...
Done! OpenMem is now available.

$ openmem-engine status
Memories: 142
  Active: 130
Edges:    87
Avg strength: 0.84

Stop losing context.

Give your agent a memory that decays, reinforces, and competes — just like yours.

Get started Star on GitHub