kunchenguid/backpass

What it solves

AI coding agents read instructions from memory files like AGENTS.md and CLAUDE.md, plus project skills, to guide their behavior. Over time these instructions become incomplete, outdated, or wrong. backpass treats these files as “weights” and uses real agent sessions as training data to propose small, evidence-backed improvements, so your agent follows what actually works in practice.

How it works

backpass scans transcript stores from seven agent harnesses (Claude, Codex, Pi, OpenCode, Grok, Cursor CLI, Hermes) directly on disk, locally or over SSH to your own machines. It associates sessions with your repository using four tiers of evidence, from exact cwd matches to recorded git remotes. It then calculates a “loss” from those sessions, aggregates feedback, and produces a set of concrete edit proposals—each add, rewrite, or remove must be supported by at least two distinct sessions with verbatim quotes. Nothing is written automatically: you run backpass to analyze, then backpass apply to review each edit and accept or reject it. One run is one bounded gradient step.

Who it’s for

Developers who use AI coding agents like Claude Code or Codex and want to systematically improve their agent’s memory files and skills based on real usage, without manually reading logs. It suits both project-level memory (one repo) and user-level memory (across all your projects on a machine).

Highlights

• Local-first: reads transcripts directly from disk, no API keys, no upload; secrets are redacted before any external agent sees them.
• Evidence-gated: every edit proposal carries verbatim quotes from at least two real sessions, so changes are grounded in actual behavior.
• Human-in-the-loop: analysis never writes; you review and decide each edit through an apply UI or terminal.
• Supports seven harnesses and SSH: includes sessions from Claude, Codex, Pi, OpenCode, Grok, Cursor CLI, and Hermes, and can pull from other machines via SSH.