01 / MEMORY
Mnemosyne remembers.
Persistent, tiered agent memory — ACT-R decay, offline consolidation, evals, 19 model backends. Context that compounds instead of resetting.
Austin Greene · Austin-rooted
I help organizations adopt AI responsibly — designing the memory, routing, governance, and workflow systems that turn models into operational capability.
Persistent, tiered agent memory — ACT-R decay, offline consolidation, evals, 19 model backends. Context that compounds instead of resetting.
Declarative, testable routing across memory, subagents, skills, and models — a PR diff for agent behavior, reviewable before deploy.
AI-ops discipline for teams running agents: token burn, model routing, guardrails, and observability — before the invoice teaches the lesson.
Where it lands in real organizations: workflow modernization, reporting architecture, incentive alignment, and responsible adoption.
memory → routing → governance, as running code
Agent memory as infrastructure: a 6-tier memory system with ACT-R decay, offline consolidation, a human-approved self-improvement loop, evals, and 19 model backends behind one API — all stdlib Python, zero runtime dependencies, 290+ tests. The memory layer of the stack.
adoption, governance & workflow rescue — where the stack deploys
A fixed-scope readiness audit for small and mid-market organizations: workflow map, token / cost-burn analysis, governance gaps, and a 90-day adoption plan. Problem → diagnosis → system design → implementation path → result.
range, shipped — exhibits, worlds, and playable systems
Twin WebGL exhibits — the Hall of the Watchers and the Hall of the Archangels — rendered in Three.js with 3D models, Doré engraving backdrops, and narrated voice-over for all 24 fallen Watchers and 12 archangels. A served, walk-through companion to the Greene Paleogenomic Museum VR.
Browser-based, StarCraft-style asymmetric RTS set in a mythic pre-flood world. Three genuinely asymmetric factions with fog of war, A* pathfinding, full economy/tech trees, and a reactive AI opponent. 100% procedural: every model, sound, and map is generated in code, zero external assets.
Isometric action-roguelike. You are the Appointed Witness, hunting the fallen Watchers and Nephilim across the Desert of Azazel. Hades-style combat with Diablo-flavored pressure and "disclosure as progression" — the loot is knowledge. Built in Godot; sprites, VFX, and music all synthesized procedurally.
smaller proofs, and the stack's ancestry
Autonomous Kalshi + Polymarket trader with cross-platform arbitrage and meta-thinking validation.
Hybrid agent with persistent memory — the ICMS/SDI foundation Mnemosyne grew from.
Agent navigator — a persona-forward front-end to the Mnemosyne / Tugboat stack. In active development.
A voluntary, rights-protected public–private model for rebuilding economic mobility in the AI era — the deployment layer, at societal scale. Policy working draft.
The Mnemosyne architecture in plain language — why memory must flow both directions, and how the dream cycle closes the loop. 14 min.
The paper — ICMS tiers, ACT-R decay, Hebbian reinforcement, and the Reflection→Instinct distillation loop. Zenodo DOI pending.
Routing as a first-class, testable object — with regret() as a "PR diff for routing." Zenodo DOI pending.
Heat-kernel coefficient for the pure-κ-jump transmission Laplacian. Targeting Letters in Mathematical Physics.
Working notes on agents, memory, trading, and applied math.
Private signal desk · research-only
Hidden operator board for Austin. Chonovisor APEX v0.3 paper-first sims. Prediction-market contracts only; not financial advice. Price-check live on Polymarket and Robinhood/Kalshi before any action.
Chonovisor: market_prior_follow. Thin edge after haircut — watch only, not a forced play. Brazil 1-1 Norway (10.4%) top simulated score.
Chonovisor: favorite_price_too_rich_pass. Model disagrees with market price — pass. Mexico 1-1 England (13.2%) top simulated score.
Paper-first discipline: market prior first, model overlay second, exact score last. Exact scores are color only, not betting anchors.
Operator rule: Chonovisor passes when model edge is negative after conservatism haircut. Watchlist entries are thin-edge monitor-only, not position signals. Sims captured 2026-07-05 10:54 local.
Open to AI operations & enablement, local-first agent systems, immersive builds, and research-to-product collaborations. Tell me what you're operating, building, or researching.