Researchers unveil MemOS, an operating system for AI memory

If you use AI daily, you know the frustration: you spend hours tuning a workflow in one chat, only for the next conversation to treat you like a complete stranger. Even worse, try taking that accumulated context over to another tool, and you are entirely back to square one.

Researchers from institutions including Shanghai Jiao Tong University and Zhejiang University unveiled MemOS, a new framework they are calling the first dedicated "memory operating system" for AI.

Treating memory like CPU cycles

Instead of treating memory as an afterthought or relying on static RAG searches, MemOS treats context as a primary computational resource that can be scheduled, shared, and evolved over time, much like an OS manages hardware.

At the heart of the system is the "MemCube"—a standardized memory unit that encapsulates everything from plain text notes to parameter-level adaptations and model activation states.

Because these memory units are standardized, they enable cross-platform memory migration. In theory, you could build customer personas in one chatbot and carry that exact context over to an automated campaign planner without missing a beat.

MemOS also uses an innovative KV-cache memory injection technique, which slashes time-to-first-token latency by up to 94% in certain setups.

Crushing benchmarks across the board

To prove their approach works, the team put MemOS through the LOCOMO benchmark, which tests memory-intensive multi-turn reasoning.

MemOS-0630 swept first place across all categories, outperforming established baselines like mem0, LangMem, Zep, and OpenAI's built-in memory.

Overall, the system beat OpenAI's memory implementation by 38.98%. In complex multi-hop and temporal reasoning scenarios—tasks that require understanding when events happened across multiple sessions—MemOS outperformed OpenAI by a massive 159%.

Why it matters

For everyday users and enterprises, the current lack of continuous memory is the single biggest barrier keeping AI agents from handling multi-week workflows. If an agent cannot remember how your team resolved a bug last Tuesday, it cannot reliably run on autopilot. MemOS provides a blueprint for persistent, reliable collaboration.

Strategically, the push for standardized memory formats is brilliant. Big platform players love closed ecosystems because your conversation history keeps you locked in. Standardizing memory into portable units threatens those walled gardens and gives control back to the user.

That said, making this truly universal will not be easy. Memory states that rely on specific model activations rarely transfer cleanly across completely different model architectures. But by framing memory management as an operating system problem rather than a search retrieval problem, MemOS shows where agent architecture needs to head next.