reflect

name	reflect
description	Process conversation history and staged items — extract facts, consolidate, then reset sessions

Reflect

Memory maintenance — extract facts from conversation history and staging inbox, then consolidate.

Phase 1: Read pending material

Three sources to check:

1. Your current session — already in your context. Extract facts from what you know.
2. Inbox sessions — list ~/.clarvis/staging/inbox/ for session_*.jsonl files. Parse each with ctools read_sessions '{"path": "<file>"}'.
3. Factoria's live session — parse ~/.clarvis/factoria/pi-session.jsonl with read_sessions (still active, not in inbox).
4. Other inbox items — check for non-session files in inbox (user-submitted summaries from /remember, staged markdown files).

If nothing new across all sources, report "nothing to reflect on" and stop.

Phase 2: Extract facts

For each piece of pending content (including your current session):

1. Scan for salient facts (people, projects, decisions, preferences, events)
2. Use recall to check if each fact already exists
3. Use remember to store genuinely new facts with appropriate type and entities

What to extract:

• People and relationships (names, roles, affiliations)
• Projects and decisions (what's being built, technical choices, progress)
• Preferences and opinions (tools, approaches, aesthetics — include confidence 0.0-1.0)
• Events and experiences (concerts, meetings, milestones)

What to skip:

• Routine tool invocations (file reads, grep, test runs)
• Ephemeral task context (open files, working directory)
• Content already in memory
• Debugging noise

Guidelines:

• fact_type: world (objective), experience (first-person), opinion (beliefs, 0.0-1.0 confidence)
• Write facts as standalone sentences — must make sense without surrounding context
• Preserve temporal anchors (dates, relative time)
• Include entity names (people, projects, places)
• Default bank: parletre. Use agora for shared knowledge.
• One fact per remember call

Phase 3: Consolidate

1. Check stats for the current state of each bank
2. Run unconsolidated to find facts not yet grouped
3. If fewer than 10 unconsolidated facts, skip consolidation
4. Use related_observations to find existing observations that cover the cluster
5. Apply consolidation decisions via consolidate

Phase 4: Mental model review

1. Check stale_models for models that need refreshing
2. Update stale models with new consolidated information

Phase 5: Complete

1. Call reflect_complete — moves inbox to staging/digested/, resets all agent sessions (current sessions move to inbox and restart fresh)
2. Report what was processed