Dreaming

How Curia’s dream engine runs background maintenance — knowledge graph decay, confidence updates, and autonomy score adjustments.

The Dream Engine is Curia’s background maintenance system, named after the neuroscience analogy: sleep is when the brain consolidates short-term experiences into long-term memory and prunes weak connections. Curia’s dream engine runs periodic passes on the knowledge graph when the system is not actively serving a conversation.

What the dream engine does

The dream engine currently runs two independent passes on configurable intervals:

Memory decay pass

Every fact and relationship in the knowledge graph has a confidence score that starts at 1.0 and decays over time based on its decay class. The decay pass applies this aging using a half-life model:

new_confidence = current_confidence × 0.5^(days_since_last_confirmed / half_life_days)

Decay class	Half-life	Effect
permanent	Never	Confidence never decreases — birth dates, legal names
slow_decay	~180 days	Halves every 6 months — employer, city of residence
fast_decay	~21 days	Halves every 3 weeks — coffee preference, current project focus

When a fact’s confidence falls below the archive threshold (default: 0.05), it is soft-deleted — archived_at is set, and the fact no longer appears in queries. The data is retained in the database for audit purposes and can be restored if later re-confirmed.

Edges follow their endpoints: when a node is archived, all edges pointing to or from it are archived in the same pass. Edges also decay and archive independently based on their own confidence.

The entire decay pass runs inside a single database transaction — partial failures leave no torn state.

Autonomy scoring pass

The scoring pass evaluates Curia’s recent actions to compute a capability score that can nudge the autonomy score up or down. It processes unscored entries in the autonomy action log:

• Deterministic scoring — approval outcomes (CEO approved or rejected an action) are scored directly
• LLM-as-judge scoring — task success and failure outcomes are evaluated by an LLM judge for quality

The composite capability score weights three dimensions:

Dimension	Weight	What it measures
Competence	45%	Task success rate, accuracy, follow-through
Commitment	35%	Reliability on recurring tasks, consistency
Compatibility	20%	Alignment with communication style and preferences

Automatic adjustments require a minimum of 30 scored actions, apply time-decay weighting (recent behavior matters more), and cannot increase the score if factual errors or overconfidence penalties are active. All changes are written to the autonomy audit record.

The decay pass and scoring pass run on independent intervals so a slow scoring pass (waiting on an LLM judge response) never blocks the decay pass from running.

When it runs

The dream engine starts automatically with Curia and runs on configurable intervals:

• Decay pass: defaults to every 24 hours (dreaming.decay.intervalMs: 86400000)
• Scoring pass: runs on its own interval, independent of the decay cadence

Both passes log their results at startup and after each run, including counts of nodes/edges decayed and archived, and pass duration.

Configuration

# config/local.yaml
dreaming:
  decay:
    intervalMs: 86400000       # how often the decay pass runs (default: daily)
    archiveThreshold: 0.05     # confidence at or below this → archived
    halfLifeDays:
      permanent: null          # never decays
      slow_decay: 180          # halves every 6 months
      fast_decay: 21           # halves every 3 weeks

See the configuration reference for the full set of options.

Re-confirmation

Decay is not a one-way street. When Curia encounters evidence that reinforces an existing fact — you mention someone’s employer in conversation, an email confirms a relationship — the fact’s last_confirmed_at timestamp updates and confidence rises. The next decay pass uses this refreshed timestamp as its reference point, effectively resetting the decay clock.

This means actively relevant facts stay fresh automatically, while facts Curia hasn’t heard about in months gracefully fade from its working knowledge.

Future passes

The dream engine is designed to host additional maintenance passes as sibling methods, each with their own config key and interval:

Pass	Purpose	Status
Memory decay	Confidence aging and archival	Implemented
Autonomy scoring	Capability evaluation and score adjustment	Implemented
Decay warning	Notify before archiving important nodes	Planned
Contradiction resolution	Detect and resolve conflicting facts	Planned
Working-memory synthesis	Consolidate conversation insights into long-term memory	Planned

Memory and the knowledge graph

How the knowledge graph stores facts, relationships, and temporal metadata.

Autonomy

The 0–100 score that the scoring pass can adjust.