Repository navigation
FAQ
No. Orrin runs fully in symbolic-only mode with an empty .env — memory, goals, control
signals, reasoning, and the UI all work. A key just unlocks the LLM tool (and SERPER_API_KEY
unlocks live web search). → Getting Started
No, and it's not claimed. Cognitive terms — "workspace," "ignition," "control signals," "attention" — name specific engineering mechanisms, not subjective experience. → What is Orrin?
The core design rule is that the LLM is the smallest part of the agent. Decision-making, goals, memory, and regulation are symbolic and inspectable; the LLM is a gated, fail-closed tool one cycle may choose to call. → Symbolic-First Design, LLM Integration
It runs a continuous cognitive loop — perceive, recall, deliberate-or-not, select, execute, learn — on its own cadence, rather than waiting for input. → The Cognitive Loop
OpenAI, Anthropic, Gemini, and any OpenAI-compatible / local endpoint, selected in the UI Settings room. The symbolic-first, fail-closed contract is identical across all of them. → LLM Integration
Yes, in a fenced way: it can write new cognitive functions into two sandboxed directories only, can't touch the selection/repair core, validates everything in a sandbox before registering, and is reviewed by the Architect peer. Review generated code before relying on it. → Self-Code and Extension
python reset_orrin.py (it snapshots first, so a reset is recoverable; --dry-run previews). A
reset rolls a new lifetime budget and starts identity/memory/learning over.
→ Existence and Lifecycle
Yes — a persistent lifetime budget (~365–730 days) is rolled on first run, counted across restarts, and eventually stops the loop, with phases that colour long-term prioritization along the way. → Existence and Lifecycle
Earlier versions churned cognition because reward was denominated in internal events. Reward is now grounded in the effect ledger — durable outward artifacts — so the incentive points at producing things. If a run looks unproductive, check that effects are being recorded. → Production and the Effect Ledger
Yes, via a tunnel or a non-loopback bind — but set ORRIN_CONTROL_TOKEN first so viewers can't
steer it. → Remote Access & Tunneling, Security Model
Configurable via ORRIN_CYCLE_SLEEP (default ~1s), with consolidation happening on longer idle
windows.
It's an experimental prototype, not security-hardened: it reads/writes local files and runs local tools. Run it on a machine you trust and don't expose control endpoints without a token. → Security Model
main.py → brain/ORRIN_loop.py → brain/loop/, then the subsystem you care about. The
Cognition Module page and docs/ARCHITECTURE.md are the maps.
- Cognitive Loop
- Workspace & Ignition
- Binding & Writeback
- Thinking & Action Selection
- Learning & Adaptation
- Effect Ledger
- Quality Standard
- Expression Membrane
- Self-Code
- Native Language Model
- Existence & Lifecycle