save

Save Session

The full transcript is saved automatically — compact captures the substance: what was decided, what went wrong, what the user wants. Not a play-by-play.

Compact is a handoff to a future agent — write what you'd need to continue this work with no other context. Be concrete and actionable, not polished.

1. Generate compact content

Two parts: a structured header for quick orientation, then a compression-gradient narrative where detail increases toward the end (recent work matters most for resumption, early work only needs enough context to understand the arc).

Skip compact for short sessions — if the compact would be comparable in size to the sum of all messages, don't generate it. Just call memory_session without the compact field; the transcript is saved automatically and is sufficient for recovery.

Goal: <why this session exists — intent, not just topic>
Status: in-progress | completed | blocked
Next: <immediate next action, one line>
Blockers: <if any, otherwise "none">

---

<Overview sentence — what was this session about, what's the outcome.>

<Early session — high compression: 1-2 sentences per topic.>

<Middle session — moderate detail: key decisions with rationale.>

<Recent work — full detail: what was tried, what worked, what didn't.
Include code snippets, exact error messages, specific function names.>

<Current state: what is working, what is broken, what is half-done.>

## Dead ends
- <approach tried → why it failed> (skip section if none)

## User requirements
- <requirements, preferences, corrections from the user>
- <specific to this project/task — things the next agent must respect>

Rules

• Substance, not mechanics — "chose X over Y because Z" matters; which files were edited does not. Focus on intent, reasoning, and constraints — things that only exist in the conversation.
• Dead ends — rejected approaches and why they were rejected, so the next agent doesn't retry them. Promote to a dedicated section.
• User requirements — preserve exact wording — paraphrase loses intent. Use direct quotes (translated to English if needed): "User: 'don't mock the database'" is better than "user prefers real DB in tests". Only quote requirements relevant to the current task — skip stale topics from earlier in the conversation that were superseded or abandoned.
• Current task in focus, history through its lens — most detail goes to what we're working on now. Earlier parts of the conversation are kept only to the extent they inform the current task: decisions still in effect, constraints discovered, context that explains why we're doing what we're doing. Drop the chronological journey — keep the causal chain.
• Length: keep total compact under 2000 characters — if space is tight, prioritize current state and recent decisions over early history. The /load budget is shared with facts and transcript tail; a bloated compact crowds them out.

2. Call memory_session mcp tool

Pass session_id: ${CLAUDE_SESSION_ID} Pass project: from the SessionStart context