Session memory extractor (Iris persona) — Hermes state.db bridge

Going-forward Iris-persona session-memory extraction. Iris is a separate Hermes profile (~/.hermes/profiles/iris/) whose conversations live in her own state.db, NOT session_*.json (write_json_snapshots off). The shared extractor's Hermes adapter only reads session_*.json, so scripts/iris_memory_capture.py is the bridge: it exports Iris's RECENT non-cron sessions (telegram + cli) from state.db to a temp dir of session_*.json, then runs hooks/session-memory-extractor.py scoped to PKA_PERSONA=iris over that dir, publishing durable memories to Vault/memory/auto/iris/.

Why the bridge, not write_json_snapshots: enabling snapshots on Iris's profile would need a gateway restart (disrupts active Iris use) and create JSON files needing pruning. The export bridge reads state.db read-only, writes temp JSON it deletes, and re-uses the extractor's .processed-ledger.json (in auto/iris/) to dedup already-seen sessions — so each run only triggers LLM calls for genuinely new conversations.

Schedule offset: 0 */4 (Luci-persona extractor is 30 */4). They share ONE global lock (PKA_ROOT/.session-extractor.lock), so the 30-min offset prevents collision; each run is bounded (--max-per-run 10, PKA_EXTRACTOR_MAX_CALLS=40) and finishes well inside the gap. Separate flock /tmp/iris-memory-extractor.lock guards the task layer.

Window: IRIS_WINDOW_HOURS=8 (overlaps the 4h cadence so nothing is missed at boundaries). The one-off full backlog is run with IRIS_BACKLOG_ALL=1 (recovered 2026-06-08). PKA_EXTRACTION_PROVIDER=agy routes extraction through Google Antigravity / Gemini quota, not Claude subscription quota; runtime_profile: direct_mixed keeps scheduler provider accounting truthful. Independent of Mission Control.