context-preservation

Context Preservation System

Long session detected or compaction imminent. Preserve critical context.

Attention Marker System

Marking Critical Information

Wrap important information that MUST survive compaction:

<!-- ATTENTION -->
[Critical information here]
<!-- /ATTENTION -->

What to Mark

Always Mark:

• Key architectural decisions
• Blocking issues and their workarounds
• Environment-specific configurations
• User preferences discovered during session
• Critical file paths or endpoints
• Authentication/connection details (sanitized)

Never Mark:

• Verbose explanations
• Intermediate debugging output
• Temporary state
• Information easily re-discovered

Example Attention Blocks

<!-- ATTENTION -->
Decision: Using PostgreSQL over SQLite for production
Reason: Need concurrent connections for API
Config: DATABASE_URL in .env.production
<!-- /ATTENTION -->

<!-- ATTENTION -->
Blocker: Tests require REDIS_URL environment variable
Workaround: export REDIS_URL=redis://localhost:6379
File: .env.test needs this added
<!-- /ATTENTION -->

Pre-Compaction Checklist

Before context is compressed, ensure:

1. All blocking issues are marked with ATTENTION
2. Key decisions are documented
3. Current task state is checkpointed
4. Critical paths/files are noted
5. Any user preferences are preserved

Checkpoint Format

Use /checkpoint command or the hook will auto-checkpoint:

## Current Task
[What we're working on]

## Completed
- [x] Step 1
- [x] Step 2

## Next Steps
- [ ] Step 3
- [ ] Step 4

## Critical Context
[ATTENTION blocks here]

Recovery After Compaction

When resuming after compaction:

1. Check .claude/memory/attention.md for preserved items
2. Check .claude/memory/checkpoint.md for task state
3. Review CLAUDE.md for current focus
4. Continue from checkpoint

Memory File Locations

File	Purpose
.claude/memory/attention.md	Preserved critical info
.claude/memory/checkpoint.md	Task state
.claude/memory/errors.md	Error patterns learned
.claude/memory/patterns.md	Behavioral patterns