Explore the architecture, models, and results. Code: github.com/MemTensor/MemT-1.… Models: huggingface.co/collections/M…

Without access to the original history at query time, the two models achieve average scores of 50.6 and 47.0 across 11 memory benchmarks, ranking #1🏅 and #2🥈 among the parametric memory methods evaluated. We also explore experience reuse in coding agents: retrieved procedural experience is written into internal memory to help the model complete subsequent tasks.

Mixture-of-Storage expands native memory into multiple experts, with separate routing for writes and reads. This open-source release includes Flash and Nano, with 288B and 36B total parameters and 17B and 3B active parameters, respectively.

During online inference, trained weights stay fixed. The persistent memory state evolves. We‘ve added native memory blocks to selected Transformer blocks, enabling the model to retain and use historical information through its internal memory state. At each turn, the model reads existing memory. Once the interaction ends, it writes the input and response into memory for future use. Throughout this process, model weights remain fixed. What updates is the memory state maintained independently for each interaction trajectory.

Can large language models retain information from past interactions and reuse it in the next task? Introducing MemT-1.0: Towards Memory-Native Evolution. This time, we've built memory directly into the model architecture. Our MemT-1.0-Nano and MemT-1.0-Flash make persistent memory part of model computation, carrying information across interactions.

A correct answer can still hide stale or unauthorized memory. Our latest article shares 12 practical measures to observe agent memory after launch: write quality, retrieval, corrections, deletion, and task outcomes. Read more to know more 👇🏻

September 23 at S-Tron Shanghai 2026 Innovation Event, our founder & CEO, Feiyu Xiong @superbearmaster, joined Angus Dunn, Patricia Bubner, PhD, and Bing Li, to discuss a question many founders are facing now: If AI lowers the barriers to building a startup, do founders still need to move to Silicon Valley? For our team, the most relevant part of the conversation was the distinction between making execution easier and making decisions better. AI can help founders code, automate operations, and iterate with smaller teams. But the harder questions remain human: Which direction is worth pursuing? How do we evaluate what we’ve built? What expertise is needed to know whether an answer is actually useful? Feiyu shared that AI-era startups will need 2 kinds of talent: practitioners who can turn AI into deliverables, and technical specialists with deep architectural and domain expertise. Tool literacy will become widespread; industry knowledge will become scarcer and more valuable. The discussion also explored why Silicon Valley’s role is changing rather than disappearing. Teams can build globally, open source can create a path to early users, and a well-positioned lightweight model can gain traction without chasing the state of the art. But biotech still needs physical labs, and strong industry networks still benefit from being in the room. Our takeaway: AI is a powerful enabler, not a substitute for judgment. Sustainable advantage still comes from strategic clarity, domain expertise, and the ability to turn insight into something people can use. Thanks to Bing, Angus, Patricia, and the S-Tron team for a thoughtful conversation. We're bringing these questions back as we continue building for an AI-native world 🪐

MemOS can help agents carry memory across sessions, so customers don't repeat their story when switching. The example follows a headphone exchange, showing reusable workflows, policy retrieval, and isolated customer memories. Check the full walkthrough👇🏻

Retrieval can’t repair a memory that was wrong when written. One lucky query becomes a reusable rule, and the error spreads. Agents need an admission layer that separates user statements, facts, inferences, and procedures first. Read more 👇🏻

Anthropic's Claude made memory visible and editable, which is real progress. But production agents need a governed memory layer: what’s stored, who can retrieve it, how it’s updated or deleted, independent of any model or chat surface. Read more 👇🏻

Agents need memory that stays useful as users, facts, and tasks change. Our latest guide explains how agent memory works with RAG and compares memory solutions on lifecycle, access control, deployment, and fit. Read the full guide 👇🏻

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…

Your agent does the work. MemOS helps it carry the context forward. 🌟 github.com/MemTensor/MemOS 📄 memos-docs.openmem.net

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.

MemOS brings relevant context back before your agent answers, then saves new questions, answers & task experience for future conversations. Less re-explaining. More continuity from one task to the next, 2 integration paths make this part of your daily workflow: ☑️ DeepSeek Harness: No-Hook AGENTS.md + Agent Skills guide DSH to retrieve memory before answering and save it after each turn. No DSH core changes or lifecycle hooks required. ☑️ Native Hook for 6 agent frameworks Codex, Claude Code, Cursor, OpenCode, OpenClaw & Hermes use framework events to trigger memory retrieval and writeback.

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.

What gets remembered? Key conclusions. Tool results. Execution traces. MemOS summarizes, indexes, and stores them locally. When a related task comes back, DSH can recover existing constraints and approaches that have already been tested—not reconstruct them from scratch. Quick Start 👇🏻

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