best sunday READ!!! the most hyped agent of the year stores memory in plain markdown and searches it with grep, @DhravyaShah took instinct apart and the moat is the harness around the files: > a profile injected at the start of every conversation: who you are, when to act vs ask, how to talk to you > every file has an id, a type, aliases for grep, and [[links]] to other files. a knowledge graph in text > the conversation reads, background commits write, consolidate and prune > old facts stay in git history unless the job decides to drop them > the cost: writes get heavier as memory grows, and stale facts pile up a $2.5b product converged on this design on its own, which makes the design the moat, and @supermemory the fastest way to ship it

I spent 3 days reverse-engineering Instinct's memory. It's beautifully simple, and works extremely well. i also wrote about how you can implement the same in supermemory in 60 lines of code!

a bad memory is a pile of chat turns ranked by similarity. the fact from march sits next to the one from today and both look equally true. so what is a good memory? - scoped to a user, project, tenant, or any entity, with reference material kept as documents in that same scope - written when it's worth remembering, so casual chatter is less likely to stick - kept as a fact that holds until updated, a preference that strengthens with repetition, or an episode that fades unless it mattered - related with updates, extends, or derives, so the latest fact wins, history stays, and a guess ranks lower until confirmed - atomic, time-aware, and willing to forget when the window closes - attributed to source documents and traceable That is supermemory.

i was rooting for V4 Pro and it came second. last month it built me a whole content dashboard, database and all just took a few tries 17% undersells it, four open models in one picker for nine dollars is a ridiculous deal either way 👀

Fun fact: In supermemory we not only learn memories, but automatically index the conversations as well. Below is an example of how @supermemory works in the Vercel @eve harness. everything works together in a harmony. Memories are formed in a coherent manner (even for things never said in a single message are learnt naturally), and the model doesn't have to choose to "remember" things. Updates, relations and forgetfulness is encoded in the data structure directly. It's multi-modal too! The context is then injected back into the model, for magically personalized and fast responses.

have been super into physical ai rabbit hole for a while.. humanoids taking over real jobs especially, and i think agi is closer than most people expect Odyssey-3 dropped today > one world model powers humanoids, robot arms, cars, drones, video games, and trains other ais > a few hours of data per task, where robots usually need thousands of hours of demos > they call it a physics agent, it speaks the language of the world and plugs straight into physical systems robots getting the shortcut humans already have is the agi that touches real jobs first.

my stack rn: - grok @bot: daily operational tasks - @claudeai: longer writing tasks - @hyperagentapp: visuals & video gen - @NousResearch hermes: long agent runs and research - @cursor_ai: anything technical i’m building all of them r wired to @supermemory so context is persistent all across my tools

my stack rn: - grok @bot: daily operational tasks - @claudeai: longer writing tasks - @hyperagentapp: visuals & video gen - @NousResearch hermes: long agent runs and research - @cursor_ai: anything technical i’m building all of them r wired to @supermemory so context is persistent all across my tools