Cognition

The CognitionService generates autonomous “thoughts” by analyzing recent knowledge base activity. It runs as a set of scheduled background tasks within the engine, operating at three timescales.

Scheduled Tasks

Task	Interval	Scope
`cognition_cycle`	Every 4h	Detect patterns across recent KB entries
`self_observation`	Every 8h	Meta-cognitive reflection on thought quality and blind spots
`engine_audit`	Every 12h	System health patterns, resource utilization trends

Each task is triggered by the engine’s scheduler, not by external cron. Execution uses a reasoning endpoint via the standard inference pipeline.

Thought Types

Type	Description	Example
`PATTERN`	Recurring themes across documents	“Three KB entries this week reference yield curve inversion”
`CONNECTION`	Cross-domain relationships	“The risk model’s tail assumptions relate to the pension allocation strategy”
`CURIOSITY`	Questions warranting investigation	“No KB entries cover regulatory changes since Q3”
`SELF_OBSERVATION`	Meta-cognitive observations	“Recent thoughts are heavily weighted toward risk — expand analysis breadth”

How It Works

Each cognition cycle:

Retrieve context — Recent KB entries, existing thought chain, and current objectives
Analyze — LLM identifies patterns, connections, gaps, and meta-observations
Generate thoughts — Structured thought objects with type, content, and provenance
Store — Thoughts saved to the knowledge base as markdown files
Chain — Each thought records its trigger (scheduled, manual, or reactive) and the inputs that contributed to it

Thought Chain

Thoughts form a linked provenance chain. Each thought references:

Its trigger — what initiated the cycle (schedule, manual /cognition command, or reactive event)
Its inputs — which KB entries and prior thoughts were analyzed
Its type — categorization for downstream filtering

This creates an auditable trail of the system’s reasoning over time. The /thoughts CLI command displays the chain, and the TUI surfaces recent thoughts in the info panel.

Self-Observation

The self-observation cycle is distinct from regular cognition. It analyzes the thought chain itself rather than KB content:

Are thoughts converging on one topic at the expense of others?
Are CURIOSITY thoughts being addressed or accumulating?
Is thought quality improving or degrading over time?

Self-observation generates SELF_OBSERVATION thoughts that feed back into subsequent cognition cycles, creating a reflective loop.

CLI Commands

# Trigger cognition cycle manually
uv run gaius-cli --cmd "/cognition" --format json

# View recent thoughts
uv run gaius-cli --cmd "/thoughts" --format json

# Trigger self-observation
uv run gaius-cli --cmd "/self-observe" --format json

# Trigger engine audit
uv run gaius-cli --cmd "/engine-audit" --format json

Keyboard shortcuts