Backpropagation for your agent's memory

What it does

backpass reads local transcripts left by seven coding-agent harnesses—including Claude, Codex, Cursor CLI, Grok, and others—and proposes edits to your AGENTS.md, CLAUDE.md, and project skills. Every suggested change must carry verbatim evidence from at least two distinct sessions, and nothing is written until you manually accept the diffs. One run equals one bounded gradient step on your agent’s memory surface.

The interesting bit

The project commits fully to the machine-learning metaphor: your memory file is weights, each session is a forward pass, the transcript is loss, and backpass literally calculates gradients to nudge the weights. It even aggregates “training data” from your other machines over SSH without installing anything remotely, piping a one-shot Node adapter through your existing ssh connection.

Key highlights

• Local-first by design: transcripts are read directly from disk, never uploaded; obvious secrets are redacted before model calls.
• Evidence-gated edits: add, rewrite, and remove operations require corroboration from at least two sessions.
• Human gate: the analysis phase never writes; only backpass apply can modify files, and it shows the evidence for each edit.
• No standalone API keys: all model calls go through acpx to a harness you have already authenticated.
• Cross-machine corpus: can collect and associate sessions from your own remote hosts via SSH, with tiered relevance scoring.

Caveats

• Requires Node >= 22.5 and acpx on your PATH; remote machines running older Node versions skip SQLite-backed harnesses.
• Windows remotes are explicitly out of scope.
• The README is detailed but truncated; some operational specifics (like exact token budget behavior) are not fully visible in the provided source.

Verdict

Worth a look if you run agent harnesses heavily and maintain a growing memory file that never seems to learn from past sessions. Skip it if you only dabble with a single agent occasionally—the evidence-gating and multi-harness setup will be overkill.