End Session

Quick Finalization Workflow

1. Collect session info: Prompt user for session summary (topic, projects touched, key activities, decisions, blockers)

2. Capture performance metrics: Get actual values automatically:

   • Run: python3 ~/.claude/get-context-metrics.py to get: tokens|raw%|ctx(u)%|overhead
   • Parse output: 95641|47.8|59.8|12.0 means:
     • Context: 95,641 tokens
     • Tokens %: 47.8%
     • Context Window %: 59.8%
     • Overhead: 12.0 percentage points

3. Generate daily log entry: Create structured log in ~/.claude-memory/providers/claude/logs/daily/YYYY.MM.DD.md with:

   • Session timestamp + topic
   • Projects touched
   • Activities realized
   • Performance metrics (see format below)
   • Next steps
   • Decisions/insights

4. Update session state: Modify ~/.claude-memory/providers/claude/session-state.md with current focus and pending tasks

5. Update integration timeline: Append to ~/.claude-memory/integration/provider-activities.md and .quick.md

6. Sync to cloud: Try pushing logs to cloud. On error, logs stay local only. (Cloud path in ~/.claude-memory/.config.json - don't read, just try sync)

7. Confirmation: Show user what was logged and sync status

Performance Metrics Format

**Context Window**: [ctx(u)]% (real usage including Anthropic overhead)
**Tokens**: [used]/200,000 ([raw]%)
**Overhead**: [difference] percentage points

Example:

**Context Window**: 59.8% (real usage)
**Tokens**: 95,641/200,000 (47.8%)
**Overhead**: 12.0 percentage points

Why both matter:

• Context Window % = What matters for hitting limits (includes system overhead)
• Tokens % = Conversation size (useful for comparing sessions)
• Overhead = Anthropic's cost (tools, prompts, cache, etc.)

Key Principle

Logs saved locally first, cloud sync is best-effort. Session finalization never blocks on cloud errors.

See existing daily logs for format examples.