@MemOS_dev
Account based inJapan!
About this account
- Account based in
- Japan
- Connected via
- China App Store
X says this location may be affected by a proxy or VPN.
Account-level information from X, not a live location or the device used for a specific post.
🧠 Memory Operating System for AI Agents 🌟 Our Repo: https://nitter.cf/t.co/PYrFDnl2Eu
𝐂𝐲𝐛𝐞𝐫 & 𝐅𝐚𝐛𝐥𝐞
Joined December 2025
- Tweets260
- Following31
- Followers381
- Likes83
Pinned Tweet
Hey guys it's been a while cuz we are having a cozy experiment
But thing's going a little bit too far…
We accidentally give a onigiri long-term memory
To make up for it we had to give this onigiri 30M tokens just to successfully se porta bonita 🍙
NOW we need somebody to adopt this cute pet
Check it out and i think u'll love Memmy
MemOS is heading to SF @Techweek_ 🚀
On October 5, we're co-hosting the Builders Beyond Borders: Vibe Coding Showcase with EchoHer, @CalHacks, and @pebble_bed in San Francisco.
Building with AI? We'd love to see what you're working on.
→ Submit your project: forms.gle/Mv1GYHJWru19y8UJA
→ Join us: partiful.com/e/mGyFYFRLUKRuP…
BREAKING NEWS: SF and LA Tech Week calendars are LIVE!
You can now register for events: tech-week.com/calendar
2,300+ event submissions across San Francisco and Los Angeles
📍SF: October 5-11
📍LA: October 12-18
People always ask about the data, so a few stats:
-SF once again takes the crown for “biggest tech week ever” with 1,500+ events
-1 in 7 SF events is about agentic AI: 213 events in SF and 53 in LA are explicitly about agents (in the title or description). In 2024 there were only about a dozen!
-There are 124 hackathons across the two weeks, and the variety is huge: everything from hardware to defense, media/creative, fintech, civic/social good, etc. Everyone is a builder now.
- LA, the OG Tech Week (now in its fifth year!), has seriously doubled down on deep tech with 100+ events and lots of El Segundo activity.
Last year, Tech Week saw attendees from over 100 countries. It’s going to be a party.
See you there! 🥳
MemOS Cloud CLI just got DSH support + Native Hook integration 🔌
Same project, new session, same background to explain again and again.
Your prefs, past decisions & lessons from completed tasks shouldn't disappear into old chats.
🤖 Made with AI
Once initialized, you can keep working while the memory flow runs with your conversations. No need to manually run memory commands each time.
⚡ On LoCoMo, MemOS CLI uses over 80% fewer tokens than OpenClaw’s native memory.
Your agent does the work. MemOS helps it carry the context forward.
🌟 github.com/MemTensor/MemOS
📄 memos-docs.openmem.net
MemOS retweeted
More like "who's not going to be there?"
Thousands of free events across SF+LA:
SF: Oct 5-11th
LA: Oct 12-18th
The official calendar launches this Tuesday!
我想很多人都有这个困扰：之前经常需要在Codex和Claude Code之间来回切换，现在又加上了最近爆火的DeepSeek Harness。
但是每次来回切换、或者想尝尝DSH的鲜的时候，我都得把背景重新讲一遍，确实挺烦的~
最近在GitHub上发现的一个开源项目 memmy-agent，把 Codex、DeepSeek Harness、Claude Code接了进去，直接解决了困扰我很久的跨Agent 记忆的问题。
我用它做的事情很简单：把分散在不同 Agent 里的历史，沉淀到同一份个人上下文里，默认 Local-first，让它们都能调用。
我把它分别接入 Codex、DeepSeek Harness、Claude Code 之后，我做了一个信号中继站小游戏。
我先在Codex 写完基础玩法，然后对话里说一句话：
第一次失败时，给玩家一次翻盘机会，别直接结束游戏。
但这句话我故意不写入代码、README 或者是本地记忆文件。
然后新开 DeepSeek Harness，它从 Memmy 直接读出了这条决定，非常丝滑。
再新开 Claude Code，它读代码前先复述了规则，再按这个方向把后续功能补完。
代码只能告诉下一个 Agent 项目做到哪，但那些留在对话里的决策，才影响着项目接下来往哪走。
以后，Agent 要记住的，不只是代码，还有你前面已经决定的决策和经验。
Memmy 不仅帮我解决了Agent记忆的问题，还让这些经验不再独属于某一个 Agent，而是只属于你。
Switch agents, not context。
Memmy 支持桌面端、CLI、API、MCP 和 Skills多种方式，可以直接接着执行任务。
如果你也经常在多个 Agent 之间切换，可以拿自己的项目试一下。
项目GitHub：github.com/MemTensor/memmy-a…
🤝 Paid partnership
🐳 @deepseek_ai Harness says: Everything is a Plugin.
We took that literally. Memory is now one too.
MemOS Local Plugin connects to a DSH profile through Cordis—no fork and no changes to DSH core. When DSH runs a task, it can retrieve context that matters instead of starting cold.
MemOS Local Plugin 2.0.16's live now, using the provider and model already set in the DSH profile. Credentials stay with DSH—no API key to copy.
🔗 memos-claw.openmem.net
🌟 github.com/MemTensor/MemOS
🎊 EXTRA EXTRA: MemTensor Named to MVP30!
Founder Park put us on MVP30 of 2026 list for our work on AI infra & long-term memory.
Once Agents live in your workflow, finishing one task isn't enough.
They must remember — and pick up exactly where they left off 🦾
A successful trajectory ≠ A successful skill
LLM agent's long-horizon runs may contain good decisions, failed attempts, retries, details that may only work in one environment, and save all of it when done.
Looks messy, and getting messier next time.
MSCE (Memory-Skill Co-Evolution) is here to organize the experience into 3 layers:
🔴 L1 keeps grounded step traces.
✨ L2 extracts reusable but revisable procedures — only becomes a skill when the evidence still supports it, the estimated gain is positive, and recent experience hasn’t contradicted it.
🟢 L3 captures what the agent has learned about the environment.
A skill still remembers where it came from, when to use it, and how to check whether it worked.
Final feedback creates another problem: it usually arrives only after a long task is over.
MSCE combines it with local self-reflections to backfill values across earlier steps — updating the memory & skills, not the base LLM weights.
🥇 MSCE on EvoAgentBench: the best or tied-best Pass@1 across all five domains.
🏆 MSCE on LoCoMo: #1 in both overall GPT-4o judge score (61.23) and F1 (49.89).
Evidence changing, skill should too.
Paper and code are here, love to hear your thought 💭
📄 Paper: arxiv.org/abs/2607.16621
🌟 Code: github.com/MemTensor/MemOS
Huge thx for Adina for sharing our work 🤗
Memory should be native to foundation models.
And that's why we release Metis to help move native memory research forward 🏃🏻➡️
Feel free to check it out 👀
Metis🚀 Memory Foundation Model released by Memtensor Research Group
Probably the first LLM that internalizes memory into its backbone👀
- 4B/9B/27B weights + code + paper
- No external RAG: memory is inside the model
- One forward pass = full read/write
- Weights frozen at inference: deploys like a standard LLM
New paper's here: Metis — the 1st prototype of a memory foundation model.
In Metis, the model weights remain fixed during inference. What changes instead is a persistent memory state.
This state carries information across interactions, updates during the forward pass, and is retrieved through memory attention when relevant — all without gradient updates.
We refer to this architecture as a memory foundation model. Metis is created to explore what becomes possible when memory is built into the model's computation, rather than added as an external system around it.
Paper, code, and checkpoints:
huggingface.co/papers/2607.2…
If this direction resonates with you, we'd appreciate a 🗳️ on @huggingface 🤗
