The technical question I kept running into was simple to state, but difficult to execute:
What happens if you design an AI system as a long-running cognitive process, with memory, audit trails, and hard safety boundaries, instead of a stateless prompt/response loop?
Most AI systems today optimize for short-term fluency. They work well in the moment, then reset. Primordia is an experiment in persistent cognition. Memory compounds over time. Reasoning can be inspected. Outputs are structurally constrained, which has tradeoffs, but avoids filtering after a response is already generated.
One clarification up front, because this always comes up. When I use the word “consciousness,” I mean it in a computational sense only: selective attention, integrated state, and metacognitive monitoring. This is not a claim about phenomenal or subjective consciousness.
Architecture (high level)
Primordia is organized as a three-layer cognitive architecture:
Layer 1: Specialized subsystems (memory, reasoning, ethics, simulation, time, world modeling) that emit typed signals rather than raw text.
Layer 2: Controllers that coordinate subsystem activity, manage arbitration, and prevent runaway behavior.
Layer 3: An integration loop inspired by Global Workspace Theory plus mandatory, fail-closed ethics enforcement.
Every response must pass ethics enforcement. Every response records which memories influenced it. Full decision provenance is stored in an append-only ledger.
Memory persists across sessions and promotes through a fixed lifecycle:
Episode → Summary → Pattern → Belief → Canon
Nothing is deleted. Every promotion retains lineage.
Some design choices
Signal integration runs in a capacity-limited workspace (50 signals max) at ~10 Hz.
Retrieval is not embedding-only. Memories are scored across significance, recency, emotional valence, access frequency, and temporal coherence.
Contradictions are first-class. Conflicting beliefs are tracked, decay without support, and must reconcile before promotion.
Ethics enforcement sits directly in the execution path. If it is unavailable, output is blocked.
What’s live
Primordia currently has 10 subsystems live, with 7 exposed through the dashboard (chat, memory, code, simulation, time, ethics, world context). All are beta-ready and actively used.
Performance: ~10s cold start, ~18–25ms per request after warm-up. Latency is higher than typical chatbots because requests route through multiple subsystems and current compute is constrained. The architecture itself is not latency-bound, but the current deployment is.
Free 3-day trial, plus a demo chat limited to 20 messages per day: https://primordiagi.com
What I’m looking for feedback on:
Does signal-based integration scale cleanly, or introduce hidden bottlenecks? What failure modes am I likely underestimating? Is append-only provenance worth the operational cost at scale, and where does it bite? Where does mandatory ethics gating break down in practice?
This is beta infrastructure, not a finished product. I’m offering founding operator access for people doing serious long-horizon work where continuity matters. The price is set to keep the group small and serious while funding ongoing infrastructure and validation work.
My working assumption is that long-horizon cognition requires structure. Whether this is the right structure is the experiment.