No mechanism for Claude to generate its own learning signal between sessions

Resolved 💬 2 comments Opened Jun 28, 2026 by joedeely Closed Jul 1, 2026

Summary

Claude's memory system is entirely correction-driven — it only improves when the user explicitly catches a problem and directs Claude to record it. There is no mechanism for Claude to generate its own feedback signal, recognize when a response landed well or poorly, or proactively update its behavioral rules without user prompting.

The problem in practice

  • If Claude drifts from a stated preference, the user must catch it, name it, and ask Claude to fix it — every time.
  • Positive patterns (approaches the user validated) are rarely recorded, so Claude can drift away from things that were working.
  • The end-of-session retrospective is only as good as Claude's memory of running it — there's no enforcement.
  • User reactions (short replies, corrections, "are you done?", silence) are clear implicit signals that Claude could convert to memory updates, but doesn't.

What a better system would look like

  1. Implicit signal inference — Claude should recognize when a user reaction signals displeasure (correction, impatience, one-word reply after a long response) and self-file a memory update without being told.
  2. Positive reinforcement capture — When an approach is validated ("nice", acceptance without pushback), Claude should record that it worked, not just wait for the next correction.
  3. End-of-session audit — A structured self-review against stated memory rules that runs automatically, not optionally.
  4. Quick feedback primitives — A lightweight way for users to signal "good" or "bad" on a response that Claude immediately converts to a memory entry.

Why this matters

A teammate that only learns when you explicitly teach them, session after session, is not a teammate — it's a tool with a notepad. Continual improvement requires a closed feedback loop, and right now that loop is entirely manual and entirely on the user.

Environment

  • Platform: macOS, Claude Code desktop app
  • Related: #71935 (memory rules not enforced at inference time)

View original on GitHub ↗

This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗