persistent memory

MemOS

Universal, local-first, persistent memory layer for AI agents. Give any LLM a memory that survives restarts — no cloud, no API keys, no vendor lock-in, and no subscription.

Quick start

terminal
$ npm install @mem-os/sdk
$ npx -y @mem-os/sdk init --yes

Key features

  • 100% local SQLite-backed storage (no cloud)
  • Graph-native with typed edges
  • Temporal validity (validFrom/validTo)
  • Trust scoring & provenance tracking
  • Confidence score state machine
  • Compact TOON context packs
  • MCP adapter (stdio + Streamable HTTP, 2026-07-28)
  • 40 MCP tools · session-start overview · task triggers
  • Review queue for inferred memories
  • Source-scoped deletion with proof-of-forgetting
  • Framework-agnostic (Ollama, LangChain, CrewAI)

Benchmarks — long-horizon recall

LoCoMo · adversarial recall @10 · higher is better
MemOS62.0%
hybrid recall + trust/provenance filters
Mem045.1%
same embedder, same protocol, infer=false

Same-protocol run: 304 LoCoMo questions, 2 conversations, identical embedder (nomic-embed-text), no reranker on either side. Overall Hit@10 was a 65.6%–65.6% tie; MemOS leads on adversarial memories by 16.9 points, where trust and provenance filtering matter. One slice, so treat it as a signal with caveats.

BenchmarkMemOSProtocol
LongMemEval-S · Hit@2099.4%500 questions · nomic 768d · saturated
LoCoMo · Hit@1065.0%1,986 questions · bge-small · repeatable
ConvoMem · Hit@1091.0%387 questions · all-MiniLM
LoCoMo same-protocol · Hit@1065.6%304 questions · vs Mem0 65.6%
End-to-end QA78.0%50 questions · independent GLM judge

Every row traces to a committed artifact or a documented run. LongMemEval retrieval is saturated at k=20 (BM25 reaches 99.6%), so we do not treat it as a differentiator. The honest comparison is the same-protocol LoCoMo slice, where the systems tie overall and separate on adversarial cases. See the methodology.

Free and open source — actually

MemOS is MIT-licensed and completely free — to use, to modify, to self-host, to ship in your product. There is no hosted platform, no usage metering, and no paid tier hiding the good features. Mem0's platform, by contrast, starts at $19/month and charges $249/month for graph memory — the kind of feature MemOS includes by default.

MemOSMem0 platform
PriceFree. Forever.Free tier capped · $19–$249/mo after
Source code100% open source (MIT)Open core — platform is proprietary
Your dataSQLite file on your diskHosted in their cloud
API keys / accountNone requiredAccount + API key required
Modify & self-hostFork it, change it, ship itNot available on the platform
Graph memoryIncluded, freePro plan ($249/mo) only

Mem0 pricing as published on mem0.ai/pricing (Hobby free with request caps, Starter $19/mo, Pro $249/mo, Enterprise custom). Mem0's open-source library exists too — but the platform most people use is the paid one, and your memories live in their cloud.