PRSM

$PRSM ca: 3jMwPagCmBfHezz6jasepJTeFiJcdEt47hJscg4epump dexscreener.com/solana/deccu… pump.fun/coin/3jMwPagCmBfHez…

Doing a live demo now! teapawt.net/i/broadcasts/2dKLlbYaN…

Q&A and Mobile App demo teapawt.net/i/broadcasts/2dKLlbYaN…

Live for the Mobile App Demo teapawt.net/i/broadcasts/1MJgNNvbW…

Another PRSM Q&A Stream tmr 7pm EST. Ben will be demoing the mobile app interface.

Going live on X and Pumpfun starting at 8 EST tonight!

I integrated github.com/safishamsi/graphi… into the agent system this week. It parses a codebase with tree-sitter and produces a full dependency graph: every import, call, and module boundary, mapped deterministically, with no LLM involved. It gives the agents grounding. When they take on a new repo, they start from the code's real structure (what actually depends on what) rather than a model's guess at how it fits together. It's a foundation, not a final answer. The graph proposes how to divide the work across teams, but the agents can still evaluate it and decide the boundaries belong elsewhere. The code supplies the facts; the agents supply the judgment. All in all, this cuts token usage not only on import but also while working by reducing the reliance on grep for codebase exploration.

2/ so I built a 3-tier recovery system. 1. mechanical reconcilers run every 5s, fix known wedges deterministically, no LLM. 2. a "doctor" agent that diagnoses the novel ones with judgment. 3. when even the doctor can't call it, it escalates to a human. the rule underneath it all: a surgeon can't operate on their own brain. the doctor is always external to what it's inspecting, and the hypervisor has zero authority to unwedge itself, only to raise its hand.

1/ spent the last 48h teaching PRSM to heal itself. PRSM runs as ONE long-lived session that never stops. great for continuity, brutal because over days it accumulates "scar damage". Agents stuck mid-task, state machines wedged in dead ends. restarting by hand just makes more scar.

12/ The long-term vision is not just a bigger model It is an adaptive cognitive architecture: a system with perception, memory, curiosity, action, prediction, and self-directed learning built into the structure itself A model that keeps learning because learning is part of how it operates

11/ From there, we can expand outward step by step After bullet hell games, we can move into top-down action games, platformers, puzzle games, survival/crafting environments, and eventually more open-ended 3D worlds Each step adds a new challenge — navigation, physics, causality, planning, memory, or long-term goals — while testing whether the same core architecture can keep adapting

10/ The first practical target is a single rich environment, likely something like bullet hell games, It gives the system perception, movement, sparse goals, open-ended exploration, planning, memory, tool use, and long-horizon learning in one testbed

9/ The action system is also part of the design Instead of only predicting text, the model should learn action spaces: moving in games, controlling simulated bodies, interacting with tools, or eventually adapting to physical systems

8/ This is where the neurotransmitter-analog system comes in Dopamine-like signals reinforce useful discoveries Serotonin-like signals regulate patience and stability Acetylcholine-like signals guide attention Norepinephrine-like signals respond to surprise and urgency

7/ Curiosity is built into the structure The model should seek useful novelty, reduce uncertainty, and learn from environments because its internal reward system values understanding It should not need every objective to be manually labeled from the outside

6/ A major focus is local learning Instead of relying only on giant centralized backpropagation runs, the architecture is designed around local predictive-coding-style updates, sparse connections, and continuous adaptation The system should learn while operating, not only during offline training

5/ Memory is split across different timescales. Short loops handle immediate perception and action Medium loops handle task progress and local reasoning Longer loops maintain goals, context, strategy, and accumulated experience The goal is hierarchical reasoning across seconds, minutes, hours, and eventually longer

4/ Then the world model learns to predict what happens next It watches sequences of states and actions, builds expectations, and updates itself when reality differs from prediction The core learning signal is prediction error: “what did I expect, and what actually happened?”

3/ The system starts with perception Raw inputs like vision, audio, proprioception, game state, or other sensor data get compressed into useful latent representations through autoencoders Those representations become the system’s internal “sensory language”