v0.2.0 — LLM Live Extraction, Speaker Attribution & Compaction

What's New in v0.2.0

🧠 LLM-Based Live Memory Extraction

The entire live fact extraction pipeline has been rebuilt. The old regex-based afterTurn() extractor has been replaced with qwen3.5:9b running locally via Ollama. This means:

• Facts are extracted with real language understanding, not pattern matching
• Technical noise (timestamps, config values, debug output) is properly filtered
• Durable facts (preferences, relationships, identity) are reliably captured

🗣️ Speaker Attribution

Previously, facts were stored as main prefers X (using the agent ID). Now:

• extractSpeakerFromText() in the JS plugin parses Discord envelope metadata to identify the real human speaker
• _resolve_speaker_name() in Python ensures facts are always stored with human names (e.g. TheDev prefers X)
• The --speaker argument is passed from JS to Python on every extraction call

🔧 Context Engine Plugin Updates

• engram_search tool — on-demand Neo4j graph memory search, registered directly in the plugin
• engram_store tool — explicit fact storage with LLM extraction fallback
• afterTurn() gate simplified — directly calls storeLiveLLM() for valid user turns, regex path fully disabled

🏥 Health Check System

New scripts/healthcheck.sh performs a 6-component automated system check:

1. Ollama status + qwen3.5:9b model loaded
2. Neo4j bolt connection
3. Graph query path (test recall)
4. LLM extraction pipeline (end-to-end smoke test)
5. Pinned facts injection
6. Graph stats (nodes, relationships, facts)

Includes opt-in ollama_auto_start config flag (default: false) — respects users who manage Ollama via systemd or Homebrew.

🧹 Cleanup

• All hardcoded paths removed — healthcheck uses dynamic $BASH_SOURCE resolution
• All personal info, agent names, and credentials removed from source
• Duplicate config.example.json removed (only config.json.example remains)
• Stale feature branch deleted

📁 Files Changed

File	Change
context_query.py	LLM extraction, speaker resolution, envelope stripping, noise filtering
session.py	Generalized help text (removed internal agent names)
scripts/healthcheck.sh	New: 6-component health check with opt-in auto-start
config.json.example	Added ollama_auto_start field and documentation
extensions/engram-context-engine/index.js	Speaker extraction, LLM-only afterTurn, engram_search/engram_store tools
extensions/engram-context-engine/manifest.json	Synced
extensions/engram-context-engine/package.json	Synced
extensions/engram-context-engine/openclaw.plugin.json	Synced

✅ Test Results

• 12/12 test suite passing (100%)
• Cross-session recall verified
• Speaker attribution verified
• Envelope noise filtering verified
• Short message gate verified
• Car insurance policy recall cross-session verified

🗜️ Session Compaction System

Engram includes a full session compaction engine that replaces OpenClaw's default compaction when enabled:

• Opt-in via ownsCompaction: true in plugin config (default: false, defers to OpenClaw runtime)
• Smart transcript splitting — keeps the last 12 messages as recent tail, compacts everything older
• Durable memory extraction — scans older messages for keywords like "remember", "favorite", "always", "never", "birthday", "policy" and extracts them before discarding
• Envelope noise filtering — strips Discord metadata, JSON blobs, tool outputs, and messages >800 chars from compaction summaries
• Structured summary generation — produces a 4-section summary: Objective, Established facts, Decisions made, Open loops
• Memory flush to disk — writes compaction summaries to memory/YYYY-MM-DD-agent-hash.md so the ingest pipeline can index them into Neo4j
• Deduplication — durable records are deduped before persisting to avoid redundant facts
• Token tracking — reports tokensBefore/tokensAfter so you can monitor compression ratios