What is a memory OS for AI agents (MemoryOS, MemOS, MIRIX)?

Short answer

A memory OS is a design in which an AI agent's memory is managed like an operating system manages computer memory: in tiers, with rules for moving, updating and retrieving what is stored. The idea started with MemGPT (2023). Best known since are MemoryOS ("Memory OS of AI Agent", 2025), MemOS from a 39-author team, and MIRIX with six memory types. All are research projects with open code.

Why “operating system”?

A computer has little fast memory and a lot of slow storage. The operating system decides what sits where and moves data back and forth, so programs behave as if there were more fast memory than there is. A language model has the same problem: a context window that is fast but limited, and everything else outside it. A memory OS applies the same discipline to the agent. It decides what stays in the window, what moves to longer-term storage, how stored items are updated and how they come back.

The first well-known paper was MemGPT, “Towards LLMs as Operating Systems” (Packer and others, October 2023). It proposed virtual context management: moving information across memory tiers to give the model extended context within its limited window, with interrupts to manage control between the model and the user. It was tested on documents larger than the window and on multi-session chat.

Four systems compared

System Paper How memory is organized Reported result
MemGPT Packer et al., 2023 Memory tiers inside and outside the context window, moved by the model Document analysis beyond the window, multi-session chat
MemoryOS Kang, Ji, Zhao and Bai, May 2025 Short-term, mid-term and long-term personal memory; modules for storage, updating, retrieval and generation On LoCoMo with GPT-4o-mini, an average of 49.11% higher F1 and 46.18% higher BLEU-1 than baselines
MemOS Li, Xi and 37 others, July 2025 MemCubes that bundle memory with metadata such as provenance and versioning, across plaintext, activation and parameter memory The abstract reports no benchmark figures
MIRIX Wang and Chen, July 2025 Six types (core, episodic, semantic, procedural, resource, knowledge vault) handled by multiple agents 85.4% on LoCoMo; on ScreenshotVQA 35% higher accuracy than a RAG baseline with 99.9% less storage

MemoryOS is the paper behind the search phrase “memory OS of AI agent”. It moves dialogue from short-term to mid-term memory first in, first out along dialogue chains, and from mid-term to long-term personal memory with a segmented page organization, a direct nod to memory paging.

MemOS is the broadest. It treats memory as a system resource with a lifecycle and includes memory that lives inside the model, such as cached activations and parameters, not only stored text. The authors describe retrieval-augmented generation as a stateless workaround without lifecycle control. The ambition is large; the abstract itself contains no measured results.

MIRIX splits memory by kind and lets separate agents manage each kind. It also ships an app that watches the screen and builds a personal memory base stored locally.

All results above are the authors’ own and come from different set-ups, so they do not rank the systems against each other. How such numbers come about is explained in How do you evaluate LLM agent memory?

The misunderstanding: a memory OS is something you install like Windows

The name suggests a platform. In practice these are research designs with open code, each with its own model of memory. What they share is a stance: memory needs management, not just a vector database. Concretely that means tiers (what is hot, what is archived), update rules (what happens when a fact changes), provenance and versions (where an entry came from) and forgetting. Those ideas are useful even if you never run any of this code. For the memory types they build on, see What are the types of memory in AI agents?

What this means for a team

The reported results do not establish how these systems handle shared team memory. A team adds problems these papers do not test: several people writing to the same memory, knowledge that must stay current for everyone, permissions, and memory that should work across tools rather than inside one agent framework. MemOS’s attention to provenance and versioning points in the right direction for that. For maintained products in this space, compare Mem0 vs Zep vs Supermemory vs Letta.

Where Bombos fits

Bombos takes the management view of memory and applies it to a team: a plugin that gives every AI tool the team uses, from Claude and ChatGPT to Copilot, Cursor, Codex and your own agents, one shared memory. It holds team knowledge, learned rules and what already sits in mail, documents and repos, and shows what an answer rests on. It runs in our cloud or on your own server with your own models, and one person can try it first for their own agents.

Related questions

What is the paper "Memory OS of AI Agent"?

It is arXiv paper 2506.06326 by Jiazheng Kang, Mingming Ji, Zhe Zhao and Ting Bai, first posted on 30 May 2025. It introduces MemoryOS, with short-term, mid-term and long-term personal memory, and the code is on GitHub at BAI-LAB/MemoryOS.

Is MemOS the same as MemoryOS?

No. MemoryOS is the three-tier system from "Memory OS of AI Agent". MemOS ("MemOS: A Memory OS for AI System") is a separate, broader proposal that also manages activation and parameter memory inside the model, not only stored text.

Should my team adopt a memory OS?

Only if you build your own agents and can maintain research code. The reported evaluations cover conversational memory and, for MemGPT, document analysis. Shared team use, permissions and governance are not what they evaluate.

Sources

  1. Memory OS of AI Agent (arXiv 2506.06326)
  2. MemOS: A Memory OS for AI System (arXiv 2507.03724)
  3. MIRIX: Multi-Agent Memory System for LLM-Based Agents (arXiv 2507.07957)
  4. MemGPT: Towards LLMs as Operating Systems (arXiv 2310.08560)
The first half hour · free

What does your AI forget?

  1. 01You tell us which AI you use and where it keeps starting over.
  2. 02We show what a shared memory changes there, on your own examples.
  3. 03You keep the plan, with or without Bombos.
Piet Baudoin
Piet BaudoinYou talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe
We reply the same working day.