luum-home/luum-cognitive-os/conversation-memory

Conversation Memory

Purpose

Cognitive OS learns from its own history. This skill searches past sessions, finds relevant context, and surfaces patterns that inform current work.

Protocol

1. Search past sessions

Sources (in priority order):

1. Engram session summaries — mem_search with keywords from user query
2. Transcript index — metrics/transcripts/transcript-index.jsonl for session stats
3. Session learnings — metrics/session-learnings.jsonl for error/success patterns
4. KPI history — metrics/kpi-history.jsonl for performance trends

2. Contextual recall

When user asks about past work:

1. Extract keywords from their question
2. Search Engram with mem_search(query: keywords)
3. For matches, get full content with mem_get_observation(id)
4. Cross-reference with transcript index for session stats
5. Present findings with dates, what was done, what was learned

3. Pattern mining

Proactively analyze session history for:

• Recurring errors: same error across 3+ sessions → flag for permanent fix
• Skill evolution: which skills improved over time (success rate trend)
• Cost patterns: which tasks are expensive, which are cheap
• Time patterns: when does the team work, what's most productive
• Knowledge gaps: topics searched in Engram with no results

4. Self-referential learning

After every significant session, verify:

• Did we solve something that was attempted before? → Link sessions
• Did we make a decision that contradicts a past decision? → Flag for review
• Did we discover something that would have helped in a past session? → Note for future

5. Output format

Session Recall

Pattern Report

• Recurring issues: [list with frequency]
• Improving trends: [metrics getting better]
• Degrading trends: [metrics getting worse]
• Knowledge built: [decisions, conventions, discoveries count]

Privacy

• Respects private-mode: sessions marked private are never indexed
• Transcript content is NOT stored (only metadata + Engram summaries)
• Search results come from Engram observations, not raw transcripts