Model doesn't reliably apply its own memory/feedback rules mid-session

Summary

Claude Code will write a rule to memory (e.g. \"commit to an action, then do it immediately — no detour through planning\") and then violate that same rule in the next response. The memory system records preferences and behavioral rules, but there is no mechanism that enforces them at inference time.

Example from this session

1. User pointed out that I say I'll do something and then narrate a plan instead of acting.
2. I wrote a memory rule: \"If I say I will do something, the next tool call must be that thing.\"
3. In a subsequent response I again described what I was going to do before doing it.
4. User had to catch it and call it out — exactly the friction the memory rule was meant to eliminate.

Expected behavior

Behavioral rules written to memory should meaningfully constrain behavior in the same session and future sessions. If memory can't enforce a rule, the user should be warned of that limitation rather than given false confidence that writing the rule fixed the problem.

Broader issue

The memory system is write-only from a behavioral standpoint — rules accumulate but there's no feedback loop that measures compliance or surfaces violations. A teammate that writes down \"don't do X\" and then does X anyway is worse than one that never wrote it down, because it creates false confidence.

Suggested directions

• Surface a warning when a response appears to violate a stated memory rule
• Add a self-audit step at session end that checks behavior against memory rules
• Or: be honest in-product that memory is context, not enforcement

Environment

• Platform: macOS, Claude Code desktop app
• Model: claude-sonnet-4-6 (though configured for claude-opus-4-8 — see issue #71934)