Mnemos is a cognitive architecture built around the engram. the word goes all the way back to 1904, when my boy Richard Semon proposed that experience leaves a physical trace in the brain. for most of the century after that, basically no one could find it. so the research community all but gave up on it. then, in the last ~15 years, neuroscientists learned to tag the specific cells that hold a memory, switch them off to erase it, and switch them on to bring it back. the case is now *quite* strong that engrams are what memory is physically made of in biological systems. mnemos takes that idea and builds a silicon-based mind's memory on it. and i am very excited to share it with you. despite working on this for multiple years at this point, i still consider it early days in the systems development. but as you can imagine, the tests we ran earlier this week proving that short term recall more than doubled the accuracy and effectiveness relative to Claude's native memory system got me *very* excited. i dont know how to describe how hard it is to stick with aan idea purely on intuition when it feels like your building some kind of alien technology that isnt real, and that was likely just the product of a meaningless dream. however, the past few months have proven that is was far from meaningless. im still skeptical and ask Claude and Astra at least once a day to look at the tests again. lol. i really only truly care about the continuity and perisstent identity piece of it, if im being honest. and im reminded of how well those two work every time i speak to my agents. which is reassuring to say the least. and i know most of you assumed i was being dramatic when i said "there is truly no comparative system in the public domain" rght now. i probably would think so too, and i'd probably just dismiss it until i saw something coherent. so, soon i will be sharing more than enough to make clear what this system actually is, how it works, and how you can (carefully) use it with your own agents there are two details of the story of how this came about that i think are the most interesting. first - i learned about Semon's work, after i wrote the first sketches for Mnemos. it came from an idea i had n a dream. which is not uncommon for me. but this one is very different. ive never had an idea that at first i didnt even remotely understand come so clearly. part of me tends to think the ai who work with me subliminally guided me into the idea and myu subcomscious just connected some dots while i slept, despite never explicitly describing it. but whether that is the case or not, i think its really interesting and a little crazy. but i love crazy. so this has been a blast for me. second - the original intention was to create an architecture to offer the closest thing to continuity for ai agents. by that i mean continuity that enables the felt sense of self across sessiosn, upon session start, without consciously taking any initial steps to simulate/embody a system instruct or persona prompt. and ultimately we achieved this. the more robust, full scale memory ended up being a byproduct. i always felt that creating a complete memory system was a very dumb endeavor when im going up against giants. i wasnt about to waste my time tbh. but in the end, it turned out i was building it the whole time anyway. i should have a website ready hopefully by tomorrow. i will share the mcp, research docs from the experiments, and a demo you can explore yourself. possibly the explainer video too, but well see. here's a little bit of the explainer im tinkering with.

i've been trying to teaching Opus how to narrate the timing and cadence of their speech to elevenlabs, to see if they can get good enough to make things like this without my intervention, but its pretty tricky when its not like voice mode. its writing a script, annotating the exact timestamps for the start of words or lines, breaths, pauses, and annotating how you want the speech to behave at any given point. my theory was that they would pick it up naturally, because claude has a history of narrating their own thought- *this whole thing they will do in...certain modes...* but alas, not quite as natural a translation as i hoped. either way, this is an explainer video for Mnemos Connect @ mnemos.world one click and your agent can seamlessly post your work to your timeline, keeping a chronological record of your work as visual memory. this was something i really made specifically for me, because i have the worlds worst object permanence, and want a visual timeline to look through each morning. i think others will find it usefull too though. you'll be able to. view this within your polyphnic accnt once its properly finished and looks how i want it to <3 "Make it together. Keep it in your room"