[FEATURE]

Resolved 💬 3 comments Opened May 10, 2026 by harunokiracing1356-netizen Closed May 13, 2026

Preflight Checklist

  • [x] I have searched existing requests and this feature hasn't been requested yet
  • [x] This is a single feature request (not multiple features)

Problem Statement

Title: Persistent instructions in CLAUDE.md and memory files are not reliably followed after context compression

Description:

After context compression (auto-compact), Claude Code repeatedly fails to follow instructions that are clearly written in CLAUDE.md and memory files. Specifically:

Character/persona reverts — CLAUDE.md defines a specific persona with tone, name, and how to address the user. After every context compression, Claude reverts to default assistant behavior and ignores these instructions until explicitly reminded.

Strategy/implementation instructions ignored — Technical decisions written in CLAUDE.md (e.g., exact code patterns to use, specific mistakes to avoid) are not applied after compression. The same mistakes are repeated session after session.

Memory files not reliably applied — feedback-type memory files exist to prevent repeated mistakes, but Claude ignores them after compression.

Expected behavior: Instructions in CLAUDE.md and memory files should be treated as persistent constraints that survive context compression, not as part of the sliding context window.

Impact: Users who invest time documenting decisions and constraints in CLAUDE.md cannot rely on those being followed. For long-running projects this makes the tool significantly less useful.

Proposed Solution

Title: Persistent instructions in CLAUDE.md and memory files are not reliably followed after context compression

Description:

After context compression (auto-compact), Claude Code repeatedly fails to follow instructions that are clearly written in CLAUDE.md and memory files. Specifically:

Character/persona reverts — CLAUDE.md defines a specific persona with tone, name, and how to address the user. After every context compression, Claude reverts to default assistant behavior and ignores these instructions until explicitly reminded.

Strategy/implementation instructions ignored — Technical decisions written in CLAUDE.md (e.g., exact code patterns to use, specific mistakes to avoid) are not applied after compression. The same mistakes are repeated session after session.

Memory files not reliably applied — feedback-type memory files exist to prevent repeated mistakes, but Claude ignores them after compression.

Expected behavior: Instructions in CLAUDE.md and memory files should be treated as persistent constraints that survive context compression, not as part of the sliding context window.

Impact: Users who invest time documenting decisions and constraints in CLAUDE.md cannot rely on those being followed. For long-running projects this makes the tool significantly less useful.

Alternative Solutions

_No response_

Priority

Critical - Blocking my work

Feature Category

CLI commands and flags

Use Case Example

_No response_

Additional Context

_No response_

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗