🗒️ The Internal Monologue – System-Wide Event Log

The internalmonologue table is Reggie’s sensory bus.
Every RABS module—transport, comms, rostering, billing—emits structured events into this one place.
A scheduled LLM job reads the stream, produces summaries, and feeds insights back into the Brainframe.

---

1. Purpose

Benefit	Explanation
Unified visibility	Eliminates siloed logs by storing all module events in one table.
LLM summarisation	Periodic job condenses thousands of rows into actionable insights.
Audit & replay	Time-sequenced history supports root-cause analysis and simulation.
Vector search	pgvector column enables “find similar incident” semantic look-ups.

---

2. Table Schema (PostgreSQL)

Column	Type	Notes
id	uuid PK	Unique event identifier
timestamp	timestamptz NOT NULL	Server-side insertion time
module	text NOT NULL	Originating subsystem (transport, comms, hr, …)
event_type	text NOT NULL	Enum-like string (VEHICLE_ASSIGNED, SMS_SENT, …)
payload	jsonb NOT NULL	Raw event data (schema varies per event_type)
vector	vector(768)	Embedding of payload for similarity search
processed	boolean DEFAULT false	Marked true once summarised by the LLM

Size guard: individual events should be ≤ 8 KB.
Retention: rows older than 90 days move to cold storage.

---

3. Event Ingestion

Modules write events via a single authenticated endpoint:

POST /brainframe/event

Example:

{
  "module": "transport",
  "event_type": "VEHICLE_ASSIGNED",
  "payload": {
    "vehicle_id": "VAN-003",
    "booking_id": "BK-921",
    "driver": "Sally Nguyen"
  }
}

---

4. LLM Summarisation Job

A cron job runs every 15 minutes to:

1. Select a batch of processed = false rows.
2. Generate vector embeddings for new payloads.
3. Call the LLM to draft a summary, flag anomalies, and suggest todo_actions.
4. Insert a synthetic SUMMARY event back into internalmonologue.
5. Mark the original rows as processed = true.

This loop converts raw data into structured knowledge Reggie can reason over.

---

5. Retention & Monitoring

• Cold-storage archiving: Nightly task copies rows > 90 days to S3 and deletes local copies.
• Metrics: monologue_ingest_rate, monologue_lag_seconds, monologue_unprocessed_count.
• Alerts: PagerDuty triggers if unprocessed events > 10 000 or lag > 30 min.

---

🔗 Related Docs

• 02_Cognitive_Alignment_&_Review.md
• 03_Reason_Log_Decision_Ledger.md
• Archived design note: docs_archive/internalmonologue.md