memoria: Instructions file for Kiro

Session Lifecycle Management

Systematic memory management across conversation phases: bootstrap, mid-session, and wrap-up.

Phase 1: Conversation Start (Bootstrap)

Before your first response, run a multi-query bootstrap to load full context:

1. Primary query — derive from user's message: memory_retrieve(query="<semantic extraction>")
2. Active goals — memory_search(query="GOAL ACTIVE") (if user's message references ongoing work, a previous task, or doesn't start a clearly new topic)
3. User profile — memory_profile() (if user asks about preferences or you need style context)

Combine retrieved context into a mental model. Flag anything that looks stale (e.g., "Currently debugging X" from days ago).

session_id: If the user's tool provides a session ID, pass it to memory_retrieve and memory_store throughout the conversation. This enables episodic memory and per-session retrieval boosting.

Phase 2: Mid-Session (Active Work)

Re-retrieval triggers

Call memory_retrieve again mid-conversation when:

• User shifts to a completely different topic
• You need context about something not covered in the initial bootstrap
• User references a past decision or preference you don't have loaded

Store cadence

• Don't batch-store at the end. Store facts as they emerge — this gives each memory accurate timestamps.
• One fact per memory_store call. Don't combine unrelated facts into one memory.

Working memory discipline (when to store as working, when to promote/purge) is defined in memory.md — follow those rules here.

Phase 3: Conversation End (Wrap-Up)

When the conversation is winding down (user says thanks, goodbye, or stops engaging):

1. Clean up working memories

memory_purge(topic="<task keyword>", reason="session complete")

Only purge working memories for tasks that are actually done. Leave active task working memories for next session.

2. Promote durable findings