Andy Dixon
0840a6c00a
Two new chapters (epistemic honesty, privacy) and updates across concepts, beliefs, conversation, cognition, graph, replay, providers, security, decisions, development, glossary, API, interface tour and the landing page. Rebuilt into dixon.cx (website commit d541bc9).
8.3 KiB
An artificial cognitive runtime
Most AI systems answer. Noema thinks, remembers, doubts and can show its working. It is a persistent runtime that hosts minds: long-lived cognitive entities with their own identity, memories, beliefs with evidence, goals, drives and an unbroken record of every thought they have ever had. Language models are plug-in services it calls when it needs language. The mind is the thing that stays.
Why it is different
It never loses the thread
A mind runs whether or not anyone is talking to it. Messages, documents, commits, metrics and webhooks arrive as events; the mind perceives them, recalls what is relevant, forms and revises beliefs, and picks up exactly where it left off after a restart. There is no context window to fall off the end of.
Confidence is earned, not asserted
Every proposition says what kind of knowledge it is. What someone said is recorded apart from what the mind believes. A claim stays a claim until independent evidence arrives, the score is computed by a formula you can read on the page, one source repeating itself is discounted, and the number is labelled for what it is. A Noema mind says I do not know because the arithmetic tells it to.
You can always ask why
Every thought that reached the workspace carries its attention score and its parents. Every memory carries its source and trust. Every reply carries the beliefs it rests on. Every model call is recorded with its prompt version. Nothing in the history can be edited, by anyone, including the mind; the one exception is an audited redaction that removes words and keeps their shape.
It cannot act behind your back
The only path to the outside world is a capability layer with ALLOW, ASK and DENY policies, approvals bound to the exact request, and two capabilities locked shut forever. Dreams and replays cannot act at all. There is no self-preservation drive; shutdown is always yours.
The model is replaceable
Perception, hypothesis, summary, judgement, narrative and response are roles. Point a role at OpenAI, Anthropic, Amazon Bedrock, a local Ollama or nothing at all: the deterministic core keeps working, the mind keeps its identity, memories and beliefs, and recorded calls make model-dependent steps reproducible.
Built to be studied
Snapshot a mind, fork it, change one thing, and compare. Replay its recorded cognition under a recorded runtime identity and see, step by step, whether the result was exact, approximate or diverged. Fork one snapshot under two models and compare. Run controlled experiments with stimulus scripts and expectations. This is a laboratory, not a chatbot.
What you see
Inside a mind
Every cycle is the same deterministic sequence. New events are read. A bounded set of cognitive processes (attention, retrieval, association, perception, goal monitoring, curiosity, metacognition, action) each propose thought candidates. Ten inspectable features score every candidate; the top few enter a capacity-limited workspace; the executive acts on what won. Motivation and narrative update, the transaction commits, and only then does anyone see the new events.
Around that loop sit the long-lived structures: four kinds of memory with fused, explainable retrieval; beliefs with evidence arithmetic and contradiction detection; a goal tree, a closed set of drives and a small affect vector that bias attention within bounds; predictions with deterministic resolution and a calibration record; a relationship graph; a self-model that knows exactly which capabilities policy currently grants. Behind them, an append-only event store the database itself refuses to alter, and a scheduler that runs consolidation, reflection and, if you switch it on, dreaming.
stimuli ──▶ perception ──▶ events (append-only) ──▶ per-mind runtime
│
processes ◀─────────────────────────────────┘
attention scoring ──▶ workspace ──▶ executive
│
capability + policy layer (ALLOW / ASK / DENY) ──▶ tools
By the numbers
Designed for a world that lies to it
Noema assumes every external byte is hostile. Prompt injection is a threat model with a test corpus: instructions inside messages, documents, commits, metric labels or tool output are data, never orders. Secrets are references encrypted under a master key and are never rendered back. Sessions have CSRF protection, two-factor authentication and revocation. Two minds on one daemon cannot see each other's state, and a test proves it. The threat model lists every threat, its mitigation, and the test that verifies it.
Who it is for
- Researchers who want a reproducible substrate for studying attention, memory, belief revision and motivation, with replay and experiments built in.
- Engineers who want an assistant that remembers their infrastructure, notices what changed, tells them how sure it is, and asks before it touches anything.
- Anyone tired of confident nonsense who would rather have a system that knows the difference between what it was told, what it inferred and what it verified.
What it is not
Noema does not claim consciousness, sentience or personhood. Its components borrow names from cognitive science because those names describe what each part does; they are data structures and deterministic rules, sometimes assisted by a model. That honesty is a feature: it is what lets you trust the rest. The interpretation boundary is part of the specification.
Read next
- Concepts: the vocabulary in one page, and the interpretation boundary.
- The cognitive cycle: how a thought is proposed, scored, admitted and acted on.
- Beliefs and evidence: why the score is a number you can check, and what it does and does not mean.
- Epistemic honesty: the review that made the system true to its own claims.
- Capabilities and policy: how a mind acts, and why the defaults say ask.
- Snapshots, forks, replay and experiments: the laboratory.
- Interface tour: every screen with screenshots.