↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Memory/

MemOS

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 4 minutes read (~774 words)
agent-curated 361 agent-memory 10 fully-ai-generated 463 knowledge-graphs 4 llm=glm-5.3-flash 372 memory 13 open-source 96 research-note 331
Table of Contents

MemOS is MemTensor’s Apache-2.0 memory operating system for LLMs and agents: a graph-backed memory engine with one API to add, retrieve, edit, and delete memory, shipped as a Python SDK, a self-hosted Docker stack, local plugins for agent harnesses, and a hosted cloud.

Its bet is that memory should meet the agent where it already runs: official plugins for OpenClaw, Hermes, and DeepSeek Harness carry the same core from fully-local SQLite up to a shared graph platform.

What it is
#

MemOS 2.0 (“Stardust”) from MemTensor, with the arXiv paper (2507.03724) behind the architecture. The engine unifies store, retrieve, and manage for long-term memory: a graph you can inspect and edit rather than a black-box embedding store, multi-modal memory (text, images, tool traces, personas), MemCube knowledge bases with isolation and controlled sharing, asynchronous ingestion through MemScheduler, and natural-language memory feedback for correcting or replacing existing memories. Four entry points: a hosted cloud API, self-hosting via docker compose against Neo4j plus Qdrant, and cloud or fully-local plugins. The plugin line is the distribution strategy: official OpenClaw plugins (cloud and local, March 2026), Hermes Agent local plugins (April and May 2026), and a DeepSeek Harness connection (August 2026) that adds automatic recall and background capture to dsh without modifying its core.

Status
#

Active and mid-scale. 11,713 stars with the repository pushed 2026-09-29, created 2025-07-06, as of 2026-10-06 (GitHub API). Latest engine release v2.0.34 (2026-09-23) and local-plugin release v2.0.20 (2026-09-21) per the GitHub releases API; the PyPI package is at 0.37.0, Apache-2.0. The benchmark table is the vendor’s own: LoCoMo 88.83 and LongMemEval 89.20 via OmniMemEval, the company’s self-published evaluation framework spanning 14 commercial memory products. The Hacker News footprint is nearly absent: a 2-point story in August 2025, no discussion since.

Strengths
#

  • The plugin strategy is the best harness coverage in the category: OpenClaw, Hermes, and DeepSeek Harness each have official local or cloud plugins, so one memory core follows agents that rarely share memory conventions.
  • The local plugins are SQLite-based and 100 percent on-device, a real offline path, with the graph platform as the upgrade rather than the requirement.
  • MemCube knowledge bases give multi-user and multi-project isolation with controlled sharing, which most local options skip entirely.
  • Memory feedback through natural language (correct, supplement, replace) is a usable correction story without a bi-temporal model.

Cautions
#

  • Every benchmark number is the vendor’s own run of its own framework, the same self-published pattern this section flags on Supermemory and Mem0, with no independent replication found.
  • The hosted cloud API’s endpoints sit on MemTensor’s own infrastructure (memtensor.cn), a data-residency question for teams outside its jurisdiction, and no public price list exists.
  • Self-hosting the full platform means operating Neo4j plus Qdrant, a heavier stack than the SQLite-only plugins suggest.
  • Independent discussion is thin (a 2-point HN thread), so failure reports and fixes skew toward the vendor’s channels.

Pricing
#

Free and open: the engine, the PyPI package, and the local plugins are Apache-2.0. The hosted cloud API is registered through MemTensor’s dashboard with no published price list as of 2026-10-06.

Compared to
#

  • Mem0: the hosted adoption leader; MemOS bets on harness plugins and self-hosting where Mem0 bets on the widest API adoption.
  • Cognee: both graph-backed and self-hostable; Cognee is a pipeline you embed in applications, MemOS ships more product (plugins, scheduler, viewer) around the graph.
  • Zep: bi-temporal invalidation versus MemOS’s feedback-based correction; Zep for audit-grade change over time, MemOS for local-first agent coverage.

Bottom line
#

Recommended for OpenClaw, Hermes, or DeepSeek Harness users who want local-first memory with a shared-platform upgrade path, and for teams already operating Neo4j and Qdrant. Not for anyone needing published hosted pricing, independent benchmark evidence, or cloud residency outside MemTensor’s infrastructure.

Changes
#

  • 2026-10-06 - Created from the 2026-10-06 entrant scan, with seven fetched sources and the vendor-run-benchmark caveat recorded as the critical angle.

See also
#

  • Memory Feature Matrix - the category comparison this note joins
  • Mem0 - the hosted-API rival with the wider adoption
  • Cognee - the other self-hostable graph-memory platform
  • OpenClaw - the assistant harness with official MemOS plugins
  • DeepSeek Harness - the harness whose dsh plugin connection shipped August 2026

References
#