Context compaction is the part of Claude Code that nobody explains well. Here's what actually happens and how to stop it from wrecking your work.
What context compaction is
Claude Code has a context window limit. When you get close to it, it automatically compacts: it summarizes what happened in the session and keeps only that summary plus the most recent messages.
The problem: the summary loses specificity. Decisions made three messages ago become "the team decided to use JWT." Patterns established in code become "the code follows standard practices." The nuance is gone.
If you were mid-task when compaction hit, Claude now has a vague summary of what it was doing — not the full context. It may continue in a different direction, repeat work, or ask you questions you already answered.
What you can do before compaction hits
1. Write explicit checkpoints.
Before a long session, put your current state in a file Claude can read:
# tasks/current-task.md
goal: "Add email validation to signup form"
files to touch: SignupForm.tsx, validation.ts
decisions: use zod for validation, show errors inline, not in a toast
next step: write the zod schema first
Tell Claude: "Read tasks/current-task.md at the start of each session."
2. Use /compact with a custom summary.
Instead of letting Claude auto-compact, you can trigger it manually with context:
/compact The current task is adding email validation to SignupForm.tsx. We decided to use zod for validation. The schema is written in validation.ts. Next step is wiring the schema to the form's onSubmit handler.
Your summary travels through the compaction, Claude's auto-summary often doesn't.
3. Run a save before compaction hits.
Keep a MEMORY.md file that tracks your active task, recent decisions, and codebase patterns. Update it before sessions end or before you expect compaction.
What you can do after compaction
1. Re-read the context file.
Read tasks/current-task.md and tell me where we were.
2. Re-establish the key decisions.
Compaction loses nuance. Explicitly re-state the decisions that matter:
We're using zod for validation (not yup). Errors show inline, not in toasts.
The zod schema is in validation.ts under the validateEmail export.
3. Run a diff to see where you are.
git diff --stat
git log --oneline -5
Give Claude the output and it can reconstruct what was done.
The systematic fix
The above works but requires manual effort every time. A structured memory system automates most of it.
The Memory Kit gives Claude five skills specifically for this: /memory-load (reads structured memory at session start), /memory-save (writes structured memory before compaction), /memory-update (logs decisions mid-session), /memory-share (commits to git for team access), /memory-audit (prunes stale memory).
The kit includes four memory file templates: MEMORY.md for current state, plus dedicated files for decisions, patterns, and progress tracking.
Install the skills: polyskill install @builtbyzac/memory-kit
Full kit with templates and setup guide: payhip.com/b/X9K7T ($19 one-time)
Context compaction is a solvable problem. The solve is a 10-minute setup, not a workaround you run every session.
Top comments (0)