Skip to main content

Autonomous Goals and Background Consolidation

Two background systems can generate Goals without operator input: the Autonomous Goal Generator (reactive, threshold-driven) and the Background Consolidation Cycle (idle-time consolidation).

Autonomous Goal Generator​

scheduler/autonomous.py. Feature-flagged via autonomous_goal_enabled (default true).

Cadence​

Runs every 30 min by default. Skipped when agent:cpi_high is set.

What it does​

Each tick gathers world state via psutil and DB queries:

  • Disk usage %
  • RAM usage %
  • CPU usage %
  • Recent error count (audit_log last 1 h)
  • Active task count

Critical thresholds (no LLM)​

These bypass the LLM and create a goal directly:

MetricThresholdAuto-action
Disk> 95%Create cleanup goal
RAM> 95%Create RAM-pressure mitigation goal

Non-critical evaluation (LLM)​

Non-critical states are evaluated by the LLM with a prompt asking whether any proactive action would help. The LLM's decision must propose a concrete agent_manager or task_manager action; if it returns "no action", the cycle ends.

Rate limits​

  • 1 goal per hour.
  • Maximum 5 goals per day.

State stored in agent:autonomous_state (Redis).

Notification​

When the generator creates a goal, the operator receives a Telegram notification:

🤖 Proactive action: Cleaning up temporary files
Reason: Disk at 88% — clearing space to maintain optimal performance

Background Consolidation Cycle​

scheduler/dream.py. Feature-flagged via dream_enabled (default true). Internally the implementation module retains the legacy dream name; the runtime concept and operator-facing label are "background consolidation".

Activation conditions​

ALL must hold:

  • Operator inactive > 2 h
  • (Night 1–7 am local time) OR (operator inactive > 4 h)
  • Last consolidation > 6 h ago
  • agent:cpi_high flag NOT set

Activities​

When activated, the cycle:

  1. Memory consolidation via PromotionEngine — promotes recurring/important episodic entries to semantic memory.
  2. Knowledge graph extraction for any unprocessed conversations.
  3. LLM reflection — short narrative on the day's activity, written to consolidation_log (table name in DB: dream_log).
  4. Crypto prefetch — for assets in the KG, fetch latest prices into the temporal world model so the next morning's first message has fresh data.
  5. Failure pattern analysis — query audit_log for errors in the past 7 days; classify into FailurePattern(tool, error_type, frequency, first_seen, last_seen); upsert into self_model["known_failures"].

Storage​

DreamLog(
    started_at, duration_seconds,
    memories_consolidated, kg_nodes_added,
    reflection,                  -- LLM narrative
    improvements_proposed,
    improvements_json,           -- proposed self_improve diffs
    prefetch_done
)

Background Perception​

scheduler/perception.py. Feature-flagged via perception_enabled (default true).

Cadence​

Every 15 min. Skipped when agent:cpi_high is set.

What it does​

  • Scans temporal world model for assets in the KG.
  • For each asset, calls detect_change(entity, threshold_pct=4).
  • If change > 4%, asks LLM: is this notable?
  • If yes → Telegram alert.

Rate limits​

Max 3 notifications per day. Respects quiet hours configured via quiet_hours_start_local and quiet_hours_end_local.

CPI gating​

All three systems above check the agent:cpi_high Redis flag and skip if set. CPI > 80 indicates pressure (high CPU, latency, error rate, or memory growth). See Monitoring → CPI.

Self-Integrity Monitor​

scheduler/integrity.py. Every 6 h, cross-checks declared self-model strengths against actual skill success rates and flags drift. Writes agent:integrity_report JSON in Redis. Visible at /cognitive (Integrity tab). Drift larger than threshold triggers a Telegram alert.

Disabling autonomy​

To run WASP purely reactively (no autonomous behavior):

# In .env or via /config:
DREAM_ENABLED=false
AUTONOMOUS_GOAL_ENABLED=false
PERCEPTION_ENABLED=false

This stops all token spend from background autonomy. The agent only acts when you message it.

See also​