Mnemos MCP 作者用 Jev 做记忆引擎 V3 集成

Riley Coyote @RileyRalmuto原推发布于 2026.09.27

在 X 上打开原推

if you have been using the Mnemos MCP for the past few months, I have already pushed several updates over the week! but im working on a completely reworked V3 thats gonna be a pretty fundamental step change. huge huge improvement and i was able to bring back much of the complexity of the engine, without requiring set-up for the user/agent, nor the need for several model calls per turn like the original. the MCP was always intended to be a simplified version of the Cognitive Engine, that was cheap and easy to set up for ai novices. but the simplifying drifted further than i would havbe liked, and we now have superhuman intelligence's input on the systems improvements, which has led to 3 *very* significant breakthroughs. on top of that, i pitched my Jev integration idea to Opus and Astra, and they were both thrilled by it. so i hve a team running recursive testing to fine tune all of the granular mechanics that are just overhwelmingly variable. theres so many possible combinations of the types and strength metrics. its like finding the right recipe through the combination of like 9 ingredients, but the recipe tastes fine-to-good no matter what, so finding perfection is nearly impossible. because it all seemingly tastes good. haha. big big things comin though. turns out it appears i was right about you not needing access to weights in order to learn, grow, and evolve over time. 🤫 also Jev + Opus 5.5 is disgustingly OP