Episodic Archiver

Analyzes conversation transcripts, embeds them for search, categorizes turns, tracks unresolved sessions for later reflection, and builds per-user behavioral profiles. Implements the self-healing agentic pattern with:

• Structured failure episodes (trigger/diagnosis/action/outcome)
• K~4 similar failure retrieval (research shows K~4 is optimal)
• Fix outcome tracking (what worked, which lessons helped)
• User behavioral profiling (communication style, expertise, bridge affinities)
• Federated taxonomy with high-fidelity LLM mode for nightly analysis

Commands

# Archive a conversation transcript
./run.sh archive transcript.json

# Archive recent sessions from all registered sources
./run.sh archive-recent --hours 24

# Deep LLM analysis of an archived session
./run.sh analyze <session_id>

# List unresolved sessions (for reflection)
./run.sh list-unresolved

# Mark a session as resolved WITH fix tracking
./run.sh resolve <session_id> --fix "What fixed it" --lessons lesson1,lesson2 --outcome success

# View fix success rate metrics
./run.sh stats

# Register a transcript source
./run.sh register <name> <path> <glob>

User Behavioral Profiling

Each session is analyzed to extract:

{
  "communication_style": "technical|casual|formal|mixed",
  "expertise_domains": ["python", "security"],
  "expertise_level": "beginner|intermediate|advanced|expert",
  "response_preferences": {
    "verbosity": "concise|balanced|detailed",
    "format": "code-first|explanation-first|mixed"
  },
  "bridge_affinities": {"Precision": 0.8, "Resilience": 0.6}
}

Profiles are incrementally merged into user_priors collection (RGMem-style):

• Bridge affinities: weighted average across sessions
• Expertise domains: union (accumulate)
• Communication style: most-recent-3-sessions voting

LLM Model Selection

• Real-time archiving: scillm quick_completion() (fast, low latency)
• Nightly analysis: deepseek-ai/DeepSeek-V3.1-TEE via CHUTES_MODEL_ID
• All LLM calls go through scillm (no raw httpx)

Storage

Collections:

• agent_conversations - Individual turns with embeddings, user_id, persona_id
• unresolved_sessions - Sessions needing follow-up (with failure episodes)
• session_summaries - LLM-analyzed session assessments with taxonomy
• user_priors - Per-user behavioral profiles (incrementally updated)

Turn categories: Task, Question, Solution, Error, Chat, Meta

Input Format

{
  "session_id": "task_123",
  "user_id": "graham",
  "persona_id": "pi",
  "messages": [
    {"from": "User", "content": "Fix the bug in auth", "timestamp": 1234567890},
    {"from": "Agent", "content": "Looking at auth.py...", "timestamp": 1234567891}
  ]
}

Integration