maxagent gives coding agents a memory of what worked last time

A developer using Claude Code or Codex solves the same problem many times, because the agent forgets every session. maxagent is an attempt to fix that.

maxagent is an open-source library on GitHub under the MIT licence. GitHub reports 265 stars in the five days since the repository was created. The project is by vzroslieigry. It works with Claude Code, Codex, Antigravity and OpenClaw.

What it does

The repository calls itself an experience base for AI coding agents. It stores short, checked recipes that the agent can look up again. The author calls them "atoms". Projects remember where work stopped. Hooks run the recall at the start of a session, and record what worked at the end.

The problem it names is familiar. An agent searches, tries a tool, fails and finds a new way. Then the next session starts from zero. maxagent tries to keep the useful parts between sessions.

What ships in the project

The project has four parts. The experience base holds the recipes. Project memory marks where a session paused and what it was doing. The hooks connect to each supported agent. A separate bridge called maxagent-bridge lets the agent drive the user's own Chrome window. It helps when a page needs a human sign-in, such as a logged-in dashboard.

The repository works across the four named agent command lines. A developer who switches between them keeps the same experience base.

How to read the traction

265 stars in five days is a real rate for a new tool. The repository has no forks yet. The project has no published issue tracker activity on its page to measure. A reader who has installed one of these tools in a hurry knows the pattern. A new agent utility collects stars fast when the shape is interesting. The useful ones are the ones that stay on the user's machine after a week. The MIT licence makes it safe to try on a work machine.

What a developer should do with it

A developer who uses Claude Code, Codex, Antigravity or OpenClaw on the same codebase every day is the target user. The bet is that recall across sessions speeds up the second and third attempt at a problem. A team that already writes a hand-maintained library of SKILL.md files should run a comparison. Check whether maxagent's atoms replace that library, or sit beside it.

The project does not report measured time savings. A developer should run a small test on a repeated task. Keep the session count, the time to first useful output, and the number of failed tool calls, before and after. That is the simplest way to see whether the memory is earning its place.

Source

• vzroslieigry/maxagent, GitHub repository.