The agent's memory is now three files on disk, and 25k stars say it works

planning-with-files is exactly what the name says: three markdown files on disk, task_plan.md, findings.md, progress.md. The agent writes its plan out, re-reads it before every decision, and updates progress as it goes. When the context window overflows or the session crashes, the plan is still sitting on disk instead of evaporating with the conversation. The trick is lifted straight from Manus's context-engineering playbook. Version 3.2.0 dropped July 3, and through the SKILL.md standard it's been picked up by 60-plus agents already, Claude Code, Cursor, Copilot, Codex, Gemini CLI. It's at 24.7k stars.

Why this matters more than it looks: goal drift is the quiet killer of long agent runs. Forty steps in, the agent has half-forgotten what it set out to do and starts diligently optimizing the wrong thing. The fix here isn't a bigger context window, it's externalizing state to files the agent re-grounds on every loop. Cheap, model-agnostic, works today, no new model required.

It pairs neatly with OpenWiki, actually. One is the repo's own knowledge, the map of what's already there. The other is the agent's working scratchpad, the memory of what it's trying to do right now. Both are the same move under the hood: stop cramming everything into the context window and put the durable stuff on disk where it survives. The filesystem quietly became the agent's long-term memory. Link: github.com/OthmanAdi/planning-with-files