PRSM

$PRSM ca: 3jMwPagCmBfHezz6jasepJTeFiJcdEt47hJscg4epump dexscreener.com/solana/deccu6e… pump.fun/coin/3jMwPagCm…

PRSM Weekly Q&A x.com/i/spaces/1kKzD…

Pushing this back to 8:30 EST, sorry for the delay

Holding a Q&A space at 8 pm EST!

This weekend, we're working on a shared memory system similar to the one in the Codex app that will allow PRSM to remember the things you've worked on and resulting lessons learned. This should be instrumental in making the PRSM system feel personal, like it learns to work with you.

5/5 Every decision is logged: inputs, scores, what got injected, and what got skipped. Why? Fully replayable.

4/5 Two slots: birth identity (cached) and runtime (refreshes as the work clarifies). So skills sharpen as the agent learns more, without thrashing the prompt cache.

1/5 Just shipped skill assignment for PRSM. Expert agents are now assigned the right skills the moment they spawn — deterministic, no model call, no vibes.

We're working on a PRSM feature called "projects" which can span multiple initiatives for larger tasks. Currently, I'm using it to build a GAN speech generator and having good success getting through initiatives without outside help.

AMA Space at 3pm today!!

Agents don’t inherently care about time so they don’t really care if training a neural network takes a few hours or a few weeks. Nor do they care about token efficiency. One of the things that we are now integrating into the system prompts is initial lightweight exploration to form a POC. After running successful smoke tests, it should dive into parallelism and hardware, acceleration for optimal, training, efficiency. It should take advantage of the resources that it has at its disposal. The POC formation is vital to avoid unnecessary token burn because you don’t want to jump into the deep end before you know you can swim. Injecting wall time into the prompts so they can more easily feel the amount of time that passing and categorize successes and failures has also been successful.

It was about 3-5x slower with no quality uplift for my application so I’ll be sticking with Gemma4. Tho in my testing I was able to get Gemma4-e4b almost on par with quality of Gemma4-26b at almost 3x speed in my mixed model test running turbo4 with better prompting so I might switch to that instead for the “heavier model”.

@bnafOg M4 Max with 36gb but I’m about to try out Qwen3.6-35b-A3b if the benchmarks are anything to go off of. Hopefully it fits. 27b is dense so it might be too slow. x.com/alibaba_qwen/s…

1) quick update on what I’m building 👇 Bradley and I are both building with PRSM to find its weak points and make it better. on my side, I’m making real progress with a personal AI assistant.

Just addressed a very nefarious bug in PRSM. We implemented a tool streaming implementation (like Claude Code) that runs tools immediately while the model streams back tokens, not waiting for the turn to finish. We also allow allowing multiple non-mutating tools to run in parallel. This is awesome for speed, but what happens if tool call A returns while tool calls B, C, D are still running. The model responds to A and changes states, modifying its tool surface interrupting the in-flight tool calls B, C, D. On replay, the provider was correctly rejecting the transcript because every assistant tool-call must have a matching tool-result. The fix was to defer refreshes until the current tool batch was fully closed and to repair old transcripts by trimming unanswered tool calls.

7) Long-term goal: as local models get better, run the full system locally on my Mac Studio (M4 Max 36GB) and eventually on a base model Mac mini. The goal is that it becomes a fully customizable personal, open source agent that runs on your desk.

6) Very much a work in progress but improving fast. Hoping to demo it soon. Also considering spinning up an AI-written blog it posts to our X so y’all can see what it’s discovering and how it works and fully open-sourcing the whole thing.