The Cognitive Loop

brain/ORRIN_loop.py is Orrin's heartbeat: a continuous cycle that runs independent of user input. Every stage is symbolic and runs with no LLM configured; the LLM is only ever an optional tool a cycle may choose to call.

flowchart LR
    P[Perceive] --> R[Recall]
    R --> W[Prepare workspace]
    W --> I{Ignition gate}
    I -- reactive --> S[Select function]
    I -- deliberate --> D[Inner loop] --> S
    S --> X[Execute]
    X --> RA[Reward accounting]
    RA --> PE[Persist]
    PE --> M[Maintain]
    M --> IC[Idle / consolidate]
    IC -.-> P

The stages

1. Perceive — gather inputs: filesystem changes, UI events, host telemetry, incoming messages.
2. Recall — retrieve relevant memories (embedding similarity + recency + strength weighting).
3. Prepare workspace — subsystems propose candidate contents; binding composes unified situations (see Binding and Workspace Writeback).
4. Ignition — the deliberation gate decides whether this cycle crosses into full deliberation or stays in low-power reactive mode (see Workspace and Ignition).
5. Select function/action — a contextual bandit picks the next cognitive function, biased by control signals, demands, cost predictions, and the workspace winner.
6. Execute — run the function; produce effects and artifacts (recorded on the effect ledger).
7. Reward accounting — assign immediate reward now and queue delayed credit for later (Learning and Adaptation).
8. Persist — write durable state and WAL checkpoints.
9. Maintain — health checks and housekeeping.
10. Idle / consolidate — at a low-power cadence, consolidate memory, replay, and account for closed time.

Cooperating daemons

The loop runs alongside off-thread daemons: the Executive advances goal steps (~7s), the memory daemon ingests/embeds/consolidates, the supervisor watches liveness and host resources, and the backend streams telemetry to the UI. See Goals: Executive vs. Daemon and Host Coupling.

The design rule

The brain never silently depends on an LLM. With no provider key configured, Orrin runs fully and simply skips LLM-backed tool calls. A non-ignited cycle stays cheap: the selector damps effortful functions (planning, codegen, research) so quiet time drifts toward low-cost work.

Tuning

• ORRIN_CYCLE_SLEEP — seconds between cycles.
• ORRIN_IGNITION_GATE, ORRIN_WORKSPACE_PRIOR, ORRIN_CONFLICT_RECRUIT — the ignition/workspace couplings (all default on, fail-safe). See the Configuration Reference.

Code pointers

• brain/ORRIN_loop.py, brain/loop/ — the loop and its phases
• Loop Phases: Detailed — a stage-by-stage walkthrough