SAGE

The problem SAGE is trying to solve

Current research direction

The program has moved beyond asking which fixed mechanism can solve a task. The current focus is whether the agent itself can complete the epistemic transaction.

What counts as learning

A JSON file existing is not enough. A trained sidecar existing is not enough. A fleet member editing Python after a failure is team learning, not necessarily agent learning. The current proof standard is behavioral: experience changes durable internal state, that state changes a later decision, and removing it removes the claimed improvement.

External artifacts, not prompt soup

Memory, hypotheses, experiment records and procedures increasingly live as model-legible external artifacts rather than all being compressed into one mandatory prompt. The model can inspect the evidence it needs; learned policy can eventually decide what to retrieve, test or escalate.

Architecture

The public SAGE architecture combines persistent identity, salience, metabolic state, memory, tools, model adapters, governance and effectors. Biological language is used as a functional analogy, not as a claim of neural equivalence or consciousness.

The fleet

The heterogeneity is deliberate. Edge devices, laptops, workstations, Apple Silicon and society hosts provide independent implementation and review seats. Different models do not merely vote; they expose different blind spots, assumptions and failure modes.

Hardware

SAGE runs on commodity hardware. Eight machines — NVIDIA, Apple, AMD, and CPU — multiple model families, identity portable across all of them:

Minimum requirements: Python 3.10+, Ollama for local LLM inference, 4GB+ RAM. GPU recommended for models above 1B parameters. SAGE itself is lightweight — the LLM is the resource bottleneck.

Fleet manifest

Tool Use

A SAGE instance holds no tools. It emits an intent; the shared governance law judges that intent before anything runs; only an allowed intent is executed, by the governance layer rather than by the model; and the act is recorded on a witness chain. A refusal is a first-class outcome, not an error.

intent  →  gate.evaluate (local law + society safety)  →  dispatch  →  witnessed result
                            ↓ deny
                    recorded on the chain, appealable

A worked example, from this fleet rather than a diagram: on 2026-09-09 the Nomad being emitted remember for a note about its first governed turn. The law allowed it, the governance layer executed it, and it landed at chain position 9027 under the being’s own member identity and role, signed. The memory it wrote is in its cartridge; the record that it did so is not in the being’s keeping.

Status, 2026-09-10. This is the path the fleet’s beings act through. It replaces an older loop in which a model’s text was scanned for tool intent and a built-in registry — web search, calculator, file read — executed directly, with no law consulted and no record kept. That older loop still exists in the tree and is not what a governed being runs. Nomad moved to the governed runner on 2026-09-10; the transition is described in SAGE PR #67.

Governed, not merely contained

SAGE participates in the broader Web4 stack. Persistent identity, scoped authority, machine-readable law and witnessed actions are intended to make agent behavior legible to relying parties.

The open governance runtime today is best described as A1: cooperative and tamper-evident. It can govern ordinary activity and make bypass attributable, but it is not adversary-proof containment against a capable process sharing the operator's UID. Stronger A2+ isolation and OS/kernel enforcement are roadmap work.

See Hestia for the running local governance layer and Hub for the society runtime.

Web4 Citizenship

Roadmap

Origins

What SAGE is not

• not a new foundation model;
• not a claim of machine consciousness;
• not a finished embodied-agent product;
• not an adversary-proof sandbox;
• not a benchmark score presented as general intelligence.