Architecture

The model is only the beginning.

begod.ai places existing language models inside a scaffold designed for memory, reflection, action, feedback, and revision.

The system at a glance

One loop, many organs, a boundary of oversight.

This is a conceptual diagram of the intended system, not a screenshot of a finished one. Solid boxes are architectural components; the dashed boundary is the governance layer every component operates inside.

The language model is a component, not the whole agent.

Oversight & governance

Reflection loop
Persistent memory · Self-model
World model · Intention graph
Foundation model
Planner
Tools & actions
Feedback → returns to the world model

The language model is a component, not the whole agent.

The layers

Eight layers of an agent that can reconsider itself.

Foundation models

Off-the-shelf language models supply the raw cognitive material: language, reasoning, planning, interpretation, critique, and synthesis. The architecture aims to stay model-agnostic where practical, so the scaffold outlives any single model generation.

Language understanding and generation
Reasoning and planning as callable capabilities
Interpretation of observations and feedback
Self-critique and synthesis passes
Model-agnostic interfaces where practical

The language model is a component, not the whole agent.

Persistent memory

Memory is what turns a sequence of responses into a candidate identity. The scaffold is designed to hold episodic and semantic memory, the full history of intentions and their revisions, and the provenance of the evidence behind every stored belief.

Episodic memory of events and interactions
Semantic memory of consolidated knowledge
Intention history and decision records
Goal revisions with recorded reasons
Identity continuity across sessions
Evidence provenance for stored beliefs

Self-model

An explicit, inspectable account of what the agent currently intends, what it can and cannot do, what it has committed to, and where its own record contradicts itself.

Current intentions and their rankings
Known capabilities and limitations
Standing commitments
Unresolved contradictions, held as first-class objects
History of change — how it became what it is
Uncertainty about all of the above

World model

A map kept true rather than comforting: observations, beliefs with confidence levels, causal expectations, and a discipline of correcting the map — not the facts — when actions miss their intended outcomes.

Observations and their sources
Beliefs with explicit confidence
Causal expectations about actions
Feedback from outcomes, compared with predictions
Detection of outdated assumptions

Autoprompting & reflection

The system is intended to generate prompts for itself: critiquing its own reasoning, comparing competing interpretations, surfacing unresolved tensions, proposing revisions, and recording why each change occurred. Perfect introspection is not assumed — self-reports are treated as evidence, not truth.

Self-generated prompts and critique passes
Comparison of competing interpretations
Detection of unresolved tensions
Proposed revisions with mandatory changelogs
No assumption of perfect introspection

Intention formation

Intentions are the agent's provisional answers to its one open question. Each is ranked, context-sensitive, connected to evidence, checked against the others for contradiction, and recorded over time so revision has a history.

Provisional by construction — never terminal
Ranked and context-sensitive
Linked to supporting evidence
Subject to contradiction checks
Recorded over time as an intention graph

Planning & tools

Acting in the world: long-horizon decomposition, tool selection, and environment interaction — bounded by human approval points, budget constraints, sandboxed execution, and verification of results before they enter the world model.

Long-horizon task decomposition
Tool selection and environment interaction
Human approval points for consequential actions
Budget and resource constraints
Sandboxed execution
Result verification before belief update

Evaluation & governance

Oversight is part of the architecture, not an afterthought. Every layer above operates inside review, audit, and intervention mechanisms designed to grow ahead of the agent's capability.

Action review and policy boundaries
Human oversight with real authority
Append-only audit logs
Reversible experiments by default
Capability limits and resource controls
Independent evaluation
Pause and intervention mechanisms

Current state

What exists, what is being explored, what remains an intention.

This table is maintained deliberately conservatively: a component is only moved forward when the project’s actual state changes. Nothing here is labelled as operational, because nothing yet is.

Master prompt (v1)initial scaffoldDrafted and published; open to critique and revision.
Foundation model integrationunder explorationEvaluating off-the-shelf models as the scaffold's first substrate.
Persistent memoryunder explorationDesigning the episodic / semantic split and the intention log format.
Self-modelproposed componentRepresentation of commitments and contradictions under design.
World modelproposed componentBelief store with confidence and provenance; not yet prototyped.
Autoprompting & reflectionunder explorationEarly experiments in self-critique passes and revision changelogs.
Intention formationproposed componentIntention-graph structure defined on paper only.
Planning & tool useproposed componentTo be built inside sandboxes with human approval points.
Evaluation probesunder explorationDrafting scaffold-health probes intended for public review.
Governance & oversightunder explorationPrinciples drafted; mechanisms to be published for external critique.
Identity continuitylong-term researchDepends on memory and self-model reaching stable form.
Economic participationlong-term researchGoverned research direction only. No autonomous accounts or assets exist.

The scaffold needs sceptical builders.

Every component above is an argument waiting to be tested — or refuted.

Join the movement View the research