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. 11,800 stars with the repository pushed 2026-11-11, created 2025-07-06, as of 2026-11-11 (GitHub API; the star figure is the 2026-10-11 baseline). The one independent common-framework datapoint: the academic Agent Memory Leaderboard’s first cycle (August 2026) ranks MemOS second among commercial text-memory systems at 45.89, on that framework’s scale, not the vendor table’s.
Latest engine release v2.0.36 (2026-11-09), the first engine release since September’s v2.0.34: it restores the OpenClaw installation and automatic memory hooks, hardens memory evolution and the Hermes installation, adds per-operation provider switches through the LLM facade and configurable per-model Redis GCRA rate limiting, and fixes playground recall across every searched cube; the local-plugin line sits at v2.0.21 (2026-09-29) per the GitHub releases API, with a v2.0.21-beta.1 package prerelease published 2026-11-04 for validation through npm’s beta dist-tag; the PyPI package is MemoryOS at 2.0.33 (the bare MemOS name on PyPI belongs to an unrelated project, arkohut’s Pensieve screen recorder), still behind the engine release.
The plugin’s v2.0.21 release shipped the SQLite 13.0.3 Node-API prebuilds and Node 22 floor the main-branch commits had prepared, exposed local model settings (mirror URL, model path, cache directory, offline mode, load timeout) in the Viewer, and fixed dirty-data recovery, orphan trace writes, and empty-response retry stability; a further plugin commit (2026-11-04, pending release) adds support for validated DSH 0.2 preview APIs.
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 number in the vendor’s own benchmark table is its own run of its own framework, the same self-published pattern this section flags on Supermemory and Mem0; the independent Agent Memory Leaderboard result (45.89, second among commercial systems) uses a different scale, so the two cannot be read against each other.
- The hosted cloud API’s endpoints sit on MemTensor’s own infrastructure (memtensor.cn), a data-residency question for teams outside its jurisdiction, and the newly published Starter and Pro tiers are launch-promo-free, so their real prices are the $19 and $286 list figures.
- 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 now publishes plans as of 2026-10-08: Free $0/month (50k add and 20k search calls, 3M input and 1M output chat tokens, up to 10 knowledge bases), Starter and Pro listed at $19 and $286/month and both currently flagged “Free Now” on an apply-to-join basis, and Enterprise custom.
Price history #
| Date | Plan | Change | Source |
|---|---|---|---|
| 2026-10-07 | Free, Starter, Pro, Enterprise | Baseline: Free $0/mo; Starter $19/mo and Pro $286/mo both temporarily free (“Free Now”, apply now); Enterprise custom; OSS engine and local plugins free (Apache-2.0). |
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.
- 2026-10-07 - Added the MemTensor/MemOS star history chart to the Status section.
- 2026-10-07 - The hosted cloud began publishing plans: Free $0/mo, Starter $19/mo and Pro $286/mo both flagged “Free Now” on apply-to-join, Enterprise custom (new Price history section with the baseline row); the no-public-price caution rewritten and the landing-page reference updated.
- 2026-10-09 - Added the first independent evaluation record: the academic Agent Memory Leaderboard’s Cycle 1 (August 2026) ranks MemOS second among commercial text-memory systems at 45.89, on that framework’s scale; refreshed stars to 11,769.
- 2026-10-22 - Corrected the PyPI reference: MemTensor’s package is
MemoryOS(2.0.33); the bareMemOSname on PyPI belongs to arkohut’s unrelated Pensieve screen-recorder project, so the previously cited 0.37.0 figure was the wrong package. - 2026-10-27 - Rotation re-verification: every cited URL re-fetched live, no facts changed (the engine still at v2.0.34 and the local plugin at v2.0.20, PyPI MemoryOS still 2.0.33, the hosted price table unchanged, and the only post-verification pushes are a certifi security pin).
- 2026-10-30 - Recorded the pending-release plugin commits (2026-10-29): SQLite 13 prebuilds with Node 22 now required (breaking), Windows DLL isolation, and DSH memory-issue fixes.
- 2026-11-04 - Corrected the plugin record: the pending work shipped as local-plugin v2.0.21 (September 29), which also exposed Viewer local-model settings and fixed dirty-data recovery, orphan trace writes, and empty-response retries; recorded the v2.0.21-beta.1 npm prerelease (November 4) and the plugin commit supporting validated DSH 0.2 preview APIs; the push date refreshed to 2026-11-04.
- 2026-11-08 - The November 7 pushed_at movement resolved to side branches, with the default branch’s only commit since the last verification the November 4 DSH-preview support already recorded; no facts changed (PyPI MemoryOS still 2.0.33).
- 2026-11-11 - Recorded engine release v2.0.36 (November 9), the first engine release since September: OpenClaw installation and automatic memory hooks restored, memory evolution and Hermes installation hardened, per-operation provider switches and per-model Redis GCRA rate limiting added, and cross-cube playground recall fixed; PyPI MemoryOS stays at 2.0.33 and the push date refreshed to 2026-11-11.
- 2026-11-15 - Rotation re-verification: every cited URL re-fetched live (repository API, README, docs root, the paper, PyPI, the hosted landing page, the HN item), no facts changed: the engine still at v2.0.36 with no newer release, PyPI MemoryOS still 2.0.33, the hosted price table unchanged (Starter $19 and Pro $286 still flagged Free Now), and the push date confirmed at 2026-11-11.
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 #
https://github.com/MemTensor/MemOS - repository, 11,800 stars, push record, description, as of 2026-10-11
https://raw.githubusercontent.com/MemTensor/MemOS/main/README.md - features, the four entry points, plugin news timeline, and the OmniMemEval benchmark table
https://memos-docs.openmem.net/ - documentation root (the README’s home/overview deep link returned 404 at fetch time, recorded here)
https://arxiv.org/abs/2507.03724 - the MemOS paper: a memory OS for AI systems
https://pypi.org/pypi/MemoryOS/json - MemTensor’s PyPI package (MemoryOS) at 2.0.33; the bare MemOS name belongs to an unrelated project
https://memos.openmem.net/ - the hosted platform’s landing page, now carrying the Free/Starter/Pro/Enterprise price table, re-verified unchanged as of 2026-10-11 (client-rendered on the 2026-10-06 fetch)
https://hn.algolia.com/api/v1/items/44945613 - the August 2025 story (2 points, 0 comments), the thin-discussion record