Backpass Turns Your AI Agent's Chat History Into Self-Improving Memory
A new CLI treats AGENTS.md as a set of weights, mining your agent transcripts to propose evidence-backed edits under a strict token budget.
PRO- Backpass treats AGENTS.md and CLAUDE.md as weights, using transcripts as a loss signal to propose evidence-backed edits.
- Install with
npm install -g backpass; requires Node 22.5 and acpx on PATH. - Reads local transcripts from seven harnesses including Claude Code, Codex, Cursor CLI, and Grok, no uploads.
- Every add, rewrite, or remove edit needs verbatim quotes from at least two distinct sessions.
- Enforces a default 5,000-token budget on always-loaded memory, with skills as overflow for narrow procedures.
- MIT licensed, source on GitHub, currently at 762 stars.
Every developer using Claude Code, Codex, or Cursor eventually ends up maintaining an AGENTS.md or CLAUDE.md file by hand, tacking on rules whenever the agent misbehaves. A new open source tool called backpass reframes that file as a set of learnable weights and mines the transcripts your agent already leaves on disk to propose concrete edits.
The pitch borrows ML vocabulary directly: your AGENTS.md is a set of weights, every agent session is a forward pass, and the transcript that session writes to disk is a loss signal that nothing currently reads. Backpass reads those transcripts, distills them, and proposes edits to your memory files. The project is MIT-licensed, written in JavaScript, and available on npm.
How the backward pass works
Installation is a single command: npm install -g backpass, requiring Node 22.5 or newer plus acpx on your PATH. The tool ships with no API keys; every model call goes through acpx to a harness you have already authenticated, so you pay through existing Claude, Codex, or Grok subscriptions rather than adding another billing surface.
The pipeline maps directly onto training loop vocabulary:
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
