`codex resume` context compression drops recent conversation context, making continuation impossible
Problem
When using codex resume to continue a previous conversation, the desktop app compresses/truncates the conversation history before loading it into a new context window. However, the compression algorithm prioritizes globally important information over recent information, which means:
- The last 3-5 turns (which contain the user's latest instructions, current task progress, recent errors, and in-progress decisions) are often discarded or severely truncated.
- Instead, the summary preserves older background context that, while important in a global sense, is useless without knowing what just happened.
- Result: the resumed session cannot continue the previous task — it's effectively amnesic.
Reproduction
- Have a long conversation (20+ turns) with concrete ongoing work (e.g., debugging a crash, configuring plugins, writing code)
- Let the session approach the context limit or naturally trigger
codex resume - Resume the conversation
- The AI agent cannot recall:
- The exact error message from the last 2 turns
- The file path or config value being edited
- The user's most recent specific instruction
- What step was next in the task
Expected Behavior
When compressing conversation history for resume, the system should use a recency-weighted approach:
| Recency | Turns | Weight | What to keep |
|---------|-------|--------|-------------|
| 🔴 Most recent | Last 3 turns | ~50% | Full user messages, errors, latest decisions |
| 🟡 Recent | Turns 4-8 | ~30% | Context, chosen approaches, file paths |
| 🟢 Mid | Turns 9-15 | ~15% | Background, failed attempts, sub-tasks |
| ⚪ Early | Older | ~5% | Project goal, completed milestones |
Minimum items any compressed summary must contain
- Current task and progress status
- User's last 3 specific requests (original text + intent)
- Most recent error messages (verbatim)
- File paths, config keys, and variable values involved
- Content explicitly marked "remember this" by the user
Impact
This makes codex resume unreliable for long-running development sessions where users need to pick up exactly where they left off.
Environment
- Codex Desktop (latest version)
- macOS
- Heavy usage: long sessions (50+ turns), multiple context-compression events per session
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗