The v2 archive recovers a candidate topology rather than a final implementation. It should be published that way.

01 – The Central Question

The key question became:

Which cognitive nodes already exist in the wild, and which actually need to be built?

That question matters because training a custom model for every cognitive role would be wasteful and probably worse than reusing strong existing specialists where they already fit.

02 – Candidate Nodes

Recovered candidate roles include:

• Router-R1: a reasoning/router component that can decide where a cognitive problem should go.
• Memory-R1: a memory-focused processor for recall, consolidation or memory judgement.
• Attention: a salience and working-memory selection processor.
• Epistemic: a processor concerned with confidence, uncertainty and knowledge gaps.
• Motivation: a processor concerned with goal pressure, unresolved tension and action-worthiness.

These names are not final product names. They are candidate functional roles.

03 – Reuse Before Training

DeepSeek R1, or an appropriate distilled R1-family model, was considered as a ready-made reasoning component. That is an important philosophical point.

The aim is not to train custom models for purity. The aim is to build the missing parts of a cognitive architecture. If an existing model already performs a role well, use it.

Custom training or adaptation should be reserved for roles where current assistant/reasoning models are structurally mismatched:

• salience;
• memory consolidation;
• recall association;
• goal selection;
• criticism;
• metacognition;
• conflict resolution;
• intent/action selection;
• affective or value modulation.

04 – Hardware Mapping

The specialist-cluster idea naturally connects to the hardware archive.

A single huge model on one machine is not the only architecture. A cluster can assign roles to available machines and GPUs:

• Hilbert for large unified-memory models;
• older high-VRAM cards for capacity-bound local inference;
• smaller cards for lightweight processors;
• discrete GPUs for image/video tasks;
• CPU or low-power systems for orchestration and state.

This is where cognitive architecture and hardware pragmatism meet.

05 – Current Status

The archive says the infrastructure/research audit should precede model training. That matters. The Lab should not imply that the cognitive cluster has already been trained if the evidence only supports topology and planning.

The current public claim is narrower and stronger: Omega is moving toward a distributed specialist-processor architecture, with existing models reused where possible and custom models reserved for genuinely missing cognitive functions.

Source note: assembled from v2 notes on the recovered candidate topology, DeepSeek R1 reuse and the correction that cluster audit/infrastructure should come before training.

The v2 archive records an important correction: an earlier Self Memory formulation should not be silently overwritten by the later procedural-memory framing. The shift is part of the research history.

01 – The Earlier Shape

The earlier memory taxonomy included:

• working memory;
• episodic memory;
• semantic memory;
• self memory.

Self Memory matters because Omega is not only trying to store facts or procedures. It is trying to preserve a continuity of identity: what the system believes about itself, its own prior states, its commitments, its habits and its recent trajectory.

That is not quite the same thing as procedural memory.

02 – The Later Shape

The later public summary uses:

• working memory;
• episodic memory;
• semantic memory;
• procedural memory.

Procedural memory is still valuable. A cognitive system needs routines, learned policies and ways of doing recurring tasks. But replacing Self Memory with Procedural Memory changes the story if it is presented as a clean upgrade.

The more honest version is that the taxonomy is being worked through.

03 – Mental Objects

Omega is better described in terms of mental objects competing for attention than scheduled model calls.

A mental object can be a goal, observation, recalled memory, hypothesis, user message, prediction error, plan, tool result or internally generated thought. The system's job is not merely to decide which timer fires next. It is to decide what deserves to occupy working memory.

This is the difference between:

• "run a thinking model every ten seconds"; and
• "when the system is idle, cognition resumes around whatever is salient."

The second version is closer to the Lab's actual target.

04 – Attention Over Scheduling

Scheduling is still needed at the engineering layer. Machines need loops, queues and execution control. But Omega's public explanation should avoid making the scheduler sound like the mind.

The cognitive model is:

• mental objects enter or leave working memory;
• activation spreads through associations;
• relevance, novelty, goals and confidence affect salience;
• weak candidates decay;
• stronger candidates compete for attention;
• selected objects guide reasoning, speech or action.

That is why "What currently occupies my mind?" is a better description than "What is in the context window?"

05 – Current Belief

The memory taxonomy should remain open in the Lab. Working, episodic, semantic and procedural memory are useful categories, but Self Memory should be preserved as a historical and possibly continuing concern.

If Omega is meant to have continuity, it needs some account of self-state. Whether that becomes its own memory class, a layer across all memory, or a set of high-priority mental objects is still a research question.

Source note: assembled from the v2 archaeological-pass correction on memory taxonomy and the recovered notes on mental objects and attention-over-scheduling.

That is a timer. It may be useful for a service loop, but it is not a satisfying model of cognition.

The preferred rule is:

As soon as idle, cognition resumes.

01 – Why The Timer Felt Wrong

A periodic timer creates the appearance of continuity while preserving the deeper assistant shape:

• wait;
• wake;
• call a model;
• produce an output;
• sleep again.

That can be engineered, but it keeps making the system feel like a queue processor wearing a cognitive mask.

Omega's goal is different. Idle time should not be empty. If no external task is active, the architecture should return to unresolved concerns, active goals, memory consolidation, prediction gaps and self-generated questions.

02 – Cognitive Transition

The recovered v2 notes point toward a cognitive-transition or control-policy model.

The question becomes:

• What state is Omega in now?
• What mental objects are currently active?
• What pressures are acting on them?
• What action, if any, should the system take next?
• Should it reason, recall, consolidate, plan, observe, speak, act or stay quiet?

This is not the same as picking the next function in a fixed pipeline. It is closer to a policy over cognitive state.

03 – Why This Matters For Will

The will problem does not vanish just because cognition runs continuously. A system can loop forever and still have no reason to prefer one concern over another.

But the control-policy framing gives the problem a place to live.

Motivation can be modelled as pressure on state transitions: unresolved goals, prediction error, novelty, repeated failure, affective value, curiosity, user commitments and self-maintenance can all affect what becomes worth acting on.

The open question is how much of that policy should be hand-built, learned, model-evaluated or split across specialist processors.

04 – Current Belief

Omega's control layer should be presented as research, not as a solved claim of machine will.

The useful public language is:

• continuous cognition;
• attention and salience;
• cognitive state transitions;
• idle thought;
• long-term concerns;
• goal pressure;
• model processors inside an architecture.

The unsafe language is anything that implies solved sentience, solved agency or a magical emergence of desire from an LLM.

Source note: assembled from the v2 archaeological-pass material on cognitive-transition/control-policy direction and the earlier "as soon as idle" design rule.