If Your Agent Forgets Everything When the Task Ends, It's a Script — Give It a Persistent Machine

Suresh935 · reddit · 2026-11-17

The author argues the spin-up-clone-run-destroy sandbox model is wrong for agents working on a project for days: quality drops before the context window fills, and compaction loses what mattered. The converging fix is treating the filesystem as memory — progress files, codebase notes, a growing AGENTS.md, git history — and giving each agent its own long-lived machine, since dependencies, warm caches, code indexes and logged-in sessions can't be rebuilt cheaply (2FA logins can't be rebuilt at all). Practices: stop idle machines instead of deleting, snapshot before risky steps, fork for parallel approaches, scope credentials per agent. Downsides — drift, stale notes, growing attack surface — call for scheduled clean-baseline rebuilds, hard idle stops, and reviewing memory files like code. Tooling notes on Fly Sprites, exe.dev, Coder and boat.dev included.

Original post →

More from coding & agent

coding & agent channel →