AI & Services

The agent that forgot yesterday's mistake, now with memory.

EverOS has been released. It is an Apache-2.0 open-source project that saves a coding agent's mistakes and context as Markdown memory for the next session.

· Huuush

The agent that forgot yesterday's mistake, now with memory

Key takeaways

  1. 1EverOS is an open-source project that records the mistakes and trial and error an agent makes and brings them back in the next session. It was updated to version 1.4.2 and is licensed under Apache-2.0.
  2. 2With one line, `uvx everos-mcp`, you can attach memory to Claude Code, Codex, Claude Desktop, and Cursor, and both EverOS Cloud and a self-hosted server are supported.
  3. 3The repository shows 30 open issues and 29 PRs, with 2 items under 'Security and quality', so it is worth checking what memory is stored and where before attaching it to a real work repository.

Would Claude Code step into the same trap it stepped into yesterday? When the session changes, the agent cannot remember that mistake. An open-source project has appeared to fill this gap. It is called EverOS, and the setup is one line.

An open-source project that helps coding agents remember their mistakes

EverOS was built to record the traps an agent falls into while working, meaning mistake patterns and trial and error, and to bring them back in the next session. It was updated to version 1.4.2 in mid-October, and the repository shows 13.5k stars and 1.0k forks. The license is Apache-2.0.

Loading X post.

Unable to load the X post. Please try again.

View original on X

The name says OS, but this is not a project for building a new agent

The repository description shows the nature of this project most briefly. "One portable memory layer for every AI agent: local-first, Markdown-native, user-owned, and self-evolving". It does not build a new agent. It separates out a memory layer that can be attached to any agent.

So what matters is where the memory lives. If memory is locked inside a specific cloud service, the moment you change tools you lose it too. EverOS aims for a memory layer that is local-first, Markdown-native, and user-owned, and it supports both EverOS Cloud and a self-hosted server. It means memory built up in Claude Code can also be used in Codex or Cursor.

There is also some background. The project was incubated by the research team of Shanda Group Investment, a Chinese investment group founded by Chen Tianqiao.

There are already quite a few cases of attaching it to several tools

The most straightforward connection is the MCP server. With one line, uvx everos-mcp, you can attach memory to Claude Code, Codex, Claude Desktop, and Cursor. It supports both EverOS Cloud and a self-hosted server you run yourself.

MCP Server for AI Coding Assistants: official integration screen connecting EverOS memory to coding tools

The plugin for Claude Code automatically saves and loads the context of past coding sessions. The structure means you do not have to explain the same things again every session.

There is also Ruminer, which is used in the browser. It is a browser extension that carries personal context across web tasks.

Browser Agent For Personal Memory: official screen of the Ruminer browser agent

Hive Orchestrator, which lets several CLI agents collaborate on one screen, and EverMem Sync, which connects memory with a single install command, are also introduced in the official repositories. On the other hand, the smart-glasses sports agent says its memory backend must be configured separately, so it is worth checking what setup is needed before you attach it.

Where it is useful for heavy users

I think there are three points in this project that are worth using. The install is one line, the memory is Markdown so you can open and edit it directly, and the memory follows you even when you change tools.

Still, it is a project that is still being refined. The repository page shows 30 open issues and 29 PRs, and the 'Security and quality' item shows 2. Before attaching it to a real work repository, it is safer to check once what memory is stored and where.

The design where you hold the memory matters more than the star count

The GitHub repository shows 13.5k stars and 1.0k forks. But a star count only shows the size of interest, and does not guarantee quality.

In my view, the default of local-first and user-owned matters more than the star number. If agent memory is tied to a particular company's server, switching tools becomes hard. EverOS has chosen a direction where the user owns that memory and can handle it as files.

What to watch next is the noise that builds up after memory accumulates

EverOS's answer to the problem of stepping into yesterday's trap again today is to save mistakes as Markdown memory so the next session reads them first. The claim that it gets smarter as memory accumulates is appealing, but old and wrong memories can pile up and become noise.

So there are two things to watch next. How quickly the open issues and PRs get cleaned up, and whether the memory really improves on its own, or whether a mechanism to filter out stale memories grows alongside it.

Reference sources

Frequently asked questions

What problem does EverOS solve?

It solves the problem of coding agents forgetting yesterday's traps when the session changes. It records mistake patterns and trial and error and brings them back in the next session.

How do I install EverOS?

If you use the MCP server, one line, uvx everos-mcp, attaches memory to Claude Code, Codex, Claude Desktop, and Cursor. Both EverOS Cloud and a self-hosted server you run yourself are supported.

Can memory built in Claude Code be used in other tools too?

Yes. The memory is structured as Markdown that the user owns, so it is introduced as usable in other tools such as Codex or Cursor. Not being locked inside a specific cloud service is the core design of this project.

Related articles