INTELLIGENCE ARCHITECT | COGNITIVE INFRASTRUCTURE FOR AGENTS
Constructing High-Dimensional Intelligence Infrastructure.
Engineering at the intersection of Epistemic Memory Systems, Agentic Context Engineering, and Governed Autonomy. Bridging the gap between retrieval and reasoning through structured doubt, externalized epistemics, and high-performance agent harnesses.
The second half of the problem: an agent that knows what it knows still needs bounds it cannot argue its way out of. Capability-locked immutable cores, policy engines as the enforcement point rather than the prompt, and self-replication that stays bounded by arithmetic instead of good intentions.
01 // SYSTEM ARCHITECTURES
- ENGRAMMIC: Epistemic memory as a graph: claims, evidence, provenance. Separates what an agent observed from what it concluded from what it invented.
- VEIL: Local-first agent memory, drop-in: sqlite-vec embeddings, FSRS decay, deterministic eviction.
- LEAP/CITE: Layered epistemic protocol with stratified types and write-time coherence enforcement.
- SKILLOPT: Text-space optimizer. Trains reusable natural-language skills for frozen agents via trajectory-driven edits and validation-gated updates.
- MONEY-MESH: Leaderless mesh of self-replicating earning agents under an immutable core. Cedar policy engine at the enforcement point, ground-truth revenue only, spend cap held by arithmetic.
- AGENT HARNESSES: Claude Code plugins, sub-agent orchestration, and slop-prevention gates across the agent lifecycle.
- ØCLOAK: At-cost RF / WiFi-sensing hardware plus a P2P threat-intel network. Surveillance commoditized; defense didn't.
02 // ONGOING EXPLORATIONS
- LeAP: Layered Epistemic Agent Protocol, stratified memory with warrant-backed claims. (shipping as Veil)
- Verification asymmetry: generation is free, verification still costs what it always did. (essay)
- Skill learning without weight updates: optimizing behaviour in text space, not parameter space.
- JEPA-native agent architectures: latent-space reasoning without text projection.
- AGM-compliant belief revision in production agent loops.
- Externalized epistemics at scale: distributed warrant chains.
- Tapestry: sovereign frontier models. Shared training, nationally-owned derivatives.
ENGRAMMIC.AI
LINKEDIN
EMAIL
KO-FI
"STRUCTURING THE UNSTRUCTURED."