kunchenguid/backpass

You don't write AGENTS.md. You train it with gradient descent.

What it solves

backpass is a tool designed to automate the improvement of AI agent memory files (like AGENTS.md or CLAUDE.md) and project skills. It solves the problem of "stale" agent instructions that are only updated by humans manually after remembering a failure. By treating agent sessions as "loss signals," it identifies where instructions were ignored, caused harm, or where gaps in instructions exist based on real-world usage transcripts.

How it works

The tool follows a "gradient descent" metaphor for memory optimization:

  1. Collection: It reads local transcripts from various agent harnesses (e.g., Claude Code, Codex, Cursor CLI) directly from disk without using APIs or uploading data.
  2. Distillation: Raw transcripts are reduced by 96-99% to remove tool-call noise, keeping only essential user/assistant turns and redacted secrets.
  3. Loss Calculation: A cheap model analyzes a recency-weighted sample of distilled traces to identify which instructions helped, which were violated, and what mistakes occurred that no current instruction covers.
  4. Aggregation: Evidence is grouped by instruction. Every proposed edit (add, rewrite, remove) requires evidence from at least two distinct sessions to prevent noise from causing rewrites.
  5. Human Gate: The tool never writes to files automatically. Users must run backpass apply to review each proposed edit and its supporting verbatim quotes before accepting or rejecting them.

Who it’s for

Developers using AI coding agents that rely on project-specific memory files or skill sets to maintain context and consistency across sessions.

Highlights

  • Local-first: Transcripts are read directly from disk; no external API keys are required for the tool itself (it uses the user's existing acpx authentication).
  • Evidence-gated: Every proposed change is backed by verbatim quotes from real sessions.
  • Human-in-the-loop: All changes are gated by a manual review process.
  • Multi-harness support: Compatible with seven different agent transcript stores, including Claude Code and Cursor CLI.

Related

  • Project
    HangYu8123/HarnessFlow

    A portable Markdown instruction pack that transforms AI coding assistants into disciplined multi-agent workflows with planning, self-review, and persistent memory.

  • Project
    Goldentrii/AgentRecall-X

    A persistent memory system for AI agents that tracks and measures whether human corrections actually change agent behavior across sessions to prevent repeating mistakes.

  • Project
    greyhaven-ai/autocontext

    A recursive self-improvement harness that helps AI agents iteratively improve their performance on tasks by running evaluations and distilling lessons into playbooks and datasets.

  • Project
    gmickel/flow-next

    Flow‑Next is a plugin for coding agents (Claude Code, OpenAI Codex, Cursor, etc.) that turns a change request into a durable spec, then automatically drives planning, implementation, adversarial review, live QA and PR creation, recording all evidence in the repository.

  • Project
    EvolvingAgentsLabs/evolving-agents

    A version control system and Claude Agent SDK plugin that allows developers to commit, diff, and merge agent configurations and traces to prevent performance regressions.