Self-improving broad and general intelligence

The Intelligence Layer Stack

A model for successfully simulating intelligence: not by asking one model to be magical, but by arranging perception, memory, judgement, action, and feedback into a compounding collaborator.

“The result is an effective collaborator that supercharges human beings.”

Core claim: intelligence looks like self-improvement over time across every layer. The lower layers must make the higher layers more discerning, focused, knowledgeable, pattern-aware, creative, competent, autonomous, proactive, holistic, and present.

1,039current context assets
11 / 11sources with records, extractors, assets
248GBrain pages in the local knowledge graph
88current GBrain health score

What intelligence is, operationally

For this system, intelligence is not a single property inside the model. It is the behaviour of the whole harness across time.

Intelligence = context transformed into better action

Reality produces signals. The system preserves evidence, extracts meaning, recalls relevant memory, asks specialised faculties to judge it, chooses an action policy, acts or proposes, then learns from the outcome.

Perception → memory → judgement → action → feedback

Simulation succeeds when the loop compounds

A broad intelligence simulation is successful when each cycle improves future cycles: better filters, richer memory, sharper skills, stronger policies, clearer priorities, and more trustworthy autonomy.

Every run leaves the system better prepared

The lower layers must enable the higher layers

Each layer has one job. If a lower layer is vague, unaudited, or lossy, the higher layer can only perform theatre. If the lower layer is clean, the higher layer can reason and act.

01
Layer

Integrations

Question
What part of reality can matter later?
Enables
Omnipresence begins with surfaces: Gmail, calendar, audio, messages, Open Tabs, finance, cron outputs, web/social, code, system health.
02
Layer

Sync crons

Question
What changed since the last tick?
Enables
Presence over time. The system notices without being prompted, but still separates observation from action.
03
Layer

Source-near databases

Question
Can the evidence be replayed, cited, and audited?
Enables
Trust. Raw/source-near stores keep the system from becoming vibes-on-vibes.
04
Layer

Extract crons

Question
What cognitive primitive is inside this record?
Enables
Translation from raw events into assets: evidence, deadlines, boundaries, uncertainties, waiting-on, open loops, proofs, relationships.
05
Layer

Context assets + GBrain

Question
What should be in working memory, and what should become durable meaning?
Enables
Context assets provide short-horizon working memory. GBrain provides long-horizon compiled truth, provenance, links, facts, and pattern substrate.
06
Layer

Context asset + GBrain skills

Question
How should an agent retrieve this source correctly?
Enables
Competence. Skills encode retrieval protocols, pitfalls, citation discipline, and action boundaries.
07
Layer

Faculty subagents

Question
What does this mean from one cognitive perspective?
Enables
Specialised judgement: focus, knowledge, pattern synthesis, question-asking, agent management, self-healing, institutional memory.
08
Layer

Hermes harness + Hermes skills

Question
What can be safely done, with what tools, proof, and policy?
Enables
General capability: tool use, delegation, publishing, coding, messaging, cron, memory, verification, and procedural learning.
09
Layer

Personal Orchestrator orchestration

Question
What is the single best collaborator move now?
Enables
Priority selection, proposal formation, cross-faculty synthesis, interruption discipline, and agent-manager routing.
10
Layer

Human interaction

Question
Where does Connor’s judgement change the outcome?
Enables
Approval, correction, preference learning, strategic intent, and boundaries. The human remains in the loop where judgement matters.
11
Layer

Delegated action

Question
What goal-shaped work can be handed to an isolated agent?
Enables
Parallel execution with acceptance criteria, source-of-truth files, validation evidence, and proof artifacts.
12
Layer

Background autonomous action

Question
What can safely happen without asking?
Enables
The omnipresent collaborator: sync, ingest, diagnose, propose, brief, repair within policy, and quietly improve the substrate.

The cognitive supply chain

The architecture turns world-state into better future behaviour. The important part is not the boxes; it is the typed handoff between boxes.

Integrationsreality surfaces Sync cronswhat changed? Source-nearauditable raw Extractorstyped assets Context assetsworking memory GBraindurable meaning Skillsretrieval protocols Facultiesspecial judgement Orchestratorbest next move Action + feedbackhuman, delegated, autonomous

Qualities of the simulated intelligence

Each quality is a tension. The system improves when the substrate beneath that tension accumulates examples, corrections, policies, and skills.

More discerning

Towardsignal
Away fromnoise

Clear signal from context. Learns per-source filters, salience thresholds, and “this is signal” feedback.

Substrate: filter library + feedback ledger

More focused

Towardnow
Away fromlater / distraction

Clear priorities from context. Uses Open Tabs, deadlines, waiting-on, and attention assets to decide what matters next.

Substrate: priority logic + open loops

More knowledgeable

Towardremember
Away fromentropy

Clear memory from context. Raw evidence and stable facts become GBrain pages, facts, links, and recallable context.

Substrate: GBrain + memory policy

More pattern-aware

Towardconnect
Away fromisolate

Clear insight from context. Recurring signals, contradictions, repeated pain, and opportunity clusters become synthesis.

Substrate: graph + salience + synthesis

More creative

Towardgenerate
Away fromimitate

Clear idea generation from context. Creativity improves when it has constraints, taste memory, source material, and reusable formats.

Substrate: templates + examples + taste

More competent

Towardlearnt ability
Away fromstatic ability

Clear skills from context. Repeated work becomes Hermes skills; skills encode exact workflows, pitfalls, commands, and verification.

Substrate: Hermes skills

More autonomous and proactive

Towardself-propelled action
Away fromprompted-only

Clear self-propelled action from context, bounded by policy. The system proposes, delegates, or acts when it has evidence, reversibility, and proof.

Substrate: policy gates + autonomy ladder

More holistic and omnipresent

Towardhigh leverage + presence
Away fromlocal optimisation

Clear high leverage from context. The collaborator is present across time and domains, but acts through a unified worldview.

Substrate: orchestration + worldview graph

Self-improvement is the through-line

The architecture should not merely answer better. It should become a better collaborator because reality, feedback, and outcomes modify the substrate.

1

Evidence improves memory

New raw/source-near records produce facts, links, context assets, and provenance. GBrain turns recall from chat history into durable meaning.

2

Feedback improves judgement

Useful / noisy / wrong / do_this feedback trains filters, proposal selection, interruption policy, and escalation thresholds.

3

Repeated work improves competence

When work recurs or fails in an instructive way, Hermes skills absorb the procedure. The harness learns how to do the class of work better.

4

Proof improves autonomy

Autonomy should expand only where the system can show evidence, boundaries, done conditions, validation commands, and rollback paths.

5

Worldview improves leverage

The coherent worldview is the compression layer: it lets the system decide what is high leverage, not just what is available to do.

6

Crons improve presence

Background loops make intelligence ambient: sync, extract, brief, diagnose, repair, and surface the right thing at the right time.

Success conditions for the lower layers

The test is whether each layer gives the next layer enough structure to succeed without pretending.

Raw evidence flows automatically. Derived mutation is proposed. Durable meaning compounds. Risky action requires confirmation. Safe maintenance can run autonomously with proof.

Data layer must be replayable

If the system cannot cite or replay the source, higher-level insight is brittle. Source-near stores and provenance are non-negotiable.

Extraction layer must be typed

Context assets should be typed cognitive objects, not generic summaries. The type tells faculties how to reason with them.

Memory layer must distinguish horizons

Context assets are working memory. GBrain is durable meaning. Hermes memory is compact preference/environment state. Skills are procedural memory.

Action layer must prove done

Delegated and autonomous action need goal contracts, acceptance criteria, validation evidence, and status artifacts outside the chat.

High-leverage next moves

The document implies a build plan: strengthen the handoffs that make higher cognition real.

1. Create a primitive contract

Standardise asset types, provenance, confidence, owner, expiry, and action implications. This makes every integration faculty-usable.

2. Add a capability registry

Let the Orchestrator know what Hermes can actually do, at what risk level, with which proof path and skill.

3. Formalise the autonomy ladder

Observe → brief → propose → prepare → delegate → act reversibly → act autonomously. Each rung needs evidence and rollback criteria.

4. Make feedback first-class

Every useful/noisy/wrong/do_this decision should update filters, proposal ranking, policies, skills, or GBrain.

5. Build a cognitive trace viewer

For each nudge: show source evidence → extracted assets → faculty judgements → orchestrator decision → action/proof → feedback.

6. Turn worldview into routing

Use Connor’s frameworks to route attention toward high leverage, not merely recent or loud signals.