Agent repeatedly ignores persistent memory instructions (reference docs, commit rules)

Summary

The agent consistently fails to follow rules stored in its own persistent memory system, despite the user investing significant effort to build up that memory. Two recurring patterns:

1. Ignores reference documentation — invents API/helper names instead of consulting documented references, even when memory explicitly instructs "re-read the project guide at the start of every session"
2. Commits and pushes without permission — memory explicitly states "never commit without explicit user request" and "run ALL tests before every commit", yet the agent routinely commits and pushes after making changes without waiting for user testing or approval

Examples

• Agent invented a non-existent template helper name instead of checking the documented helper API table — took 4 correction rounds to resolve what was a single-line change
• Agent ran git commit && git push immediately after a code fix, without the user testing or saying "check in please" — despite memory rules explicitly prohibiting this (happened multiple times in a single session)

Expected behaviour

When the agent has persistent memory rules (via memory files or repeated instruction), it should follow them reliably. The user should not need to re-correct the same behaviours across multiple sessions.

Impact

The user invested significant effort training the agent via memory files, a project guide, and repeated feedback. The agent ignoring these resources undermines that investment and erodes trust. The user reports this happens "at least 3 times a day".

Environment

• Claude Code (VS Code extension)
• Model: Claude Opus 4.6
• Memory system: active with ~20 memory files covering user preferences, feedback rules, project context