ONA is a lightweight, real-time implementation of NARS created by Dr. Patrick Hammer. ONA is not installed by default in OmegaClaw. There are no ONA bindings in the repository, no lib_ona.metta, and no dedicated operator in the (metta ...) skill surface. The stock install ships two formal engines — symbolic NAL (|-) and probabilistic PLN (|~).
This page exists because ONA is named in the project's architectural vision and has been an experimental target — see Status below.
- Throughput — ONA can process thousands of inference steps per second.
- Temporal reasoning — reasoning about before/after, cause-and-effect, and sensorimotor loops.
- Real-time capability — inference on a millisecond budget.
If ONA were installed, the following shapes would route to it:
| Situation | Engine (if installed) |
|---|---|
Known chain A → B → C, static facts |
NAL |- |
| Property-based categorical inference | PLN |~ |
| Event sequences, cause/effect from experience | ONA |
| Real-time sensorimotor loops | ONA |
Today, with ONA absent, the current practical fallback for temporal questions is NAL plus external temporal grounding — have the agent fetch timestamps or event ordering through (shell ...) or (search ...) and atomize them into NAL premises. This is not equivalent to what ONA would provide, and the docs do not pretend otherwise.
There is no stock invocation surface for ONA in OmegaClaw. lib_omegaclaw.metta does not load ONA bindings. The (metta ...) skill has no ONA operator.
Any integration has to be built — either by calling an external ONA binary through (shell ...) and parsing its output, or by constructing MeTTa-side wrappers that emulate ONA's temporal operators. This is a non-trivial tools-authoring task and is the subject of the experiment described below.
OmegaClaw's architectural vision lists ONA as a third engine, but in practice ONA support has been pursued as an experiment in whether the agent can itself build the tools to interface with that reasoning system. Results so far: partly successful.
What this means concretely:
- The agent has been used to draft shell-callable wrappers and MeTTa-side glue code for ONA-style inference. Some components work in isolation; none are integrated into the default agent loop.
- Temporal inference is not meaningfully exposed in the current codebase.
- There is no native mechanism for belief revision triggered by time — confidence does not automatically decay for stale claims.
- Real-time feedback loops require careful phase tracking and frequently break due to state-management failures when the agent attempts to orchestrate them end-to-end.
Treat ONA in OmegaClaw as a research direction, not a feature. The limits of agent-authored self-improvement (which is effectively what this experiment is) are documented in reference-failure-modes.md §6 — self-authored improvements should be taken as signals but validated externally before being relied upon.
- reference-lib-nal.md — the symbolic engine (installed).
- reference-lib-pln.md — the probabilistic engine (installed).
- reference-orchestration.md — how the LLM picks between the installed engines.
- reference-failure-modes.md — §6 covers the self-improvement pattern this experiment fits into.