John Martin

Carpenter building open-source AI and crypto by directing AI. Lu: an AI that says "I don't know." Cattledog: honesty for agents. PQ VPN in progress.

[John Martin @dialectforge]

worked on my pc with grok build, then picked up my iphone and asked claude and grok what we just did. both knew. same memory, two different ais.

that's FlowState. free, open source, runs on your own machine. just shipped v2.1.0

https://t.co/daB3boaoin

[John Martin @dialectforge]

@ColateAi yeah the why is the part that saves you later. i ended up building my own for it, local and free so it works across whatever model i'm using
https://t.co/daB3boaoin

[John Martin @dialectforge]

@HaStrive correction on my end, right now any agent can relabel a guess. making that user-only is next on my list, your point pushed it up

[John Martin @dialectforge]

@HaStrive I agree, in-place rewrites are what killed symlinks for me too. the tagging ended up being the part i use most, when something looks off i can see which agent wrote it and when

[John Martin @dialectforge]

@098uyiubgvgcss same conclusion here. flowstate flags the conflict and keeps both, tagged with which app and model wrote each. working on 2.4 now so the agent has to stop and get your ok before anything gets overwritten. latest-wins is how bad guesses win quietly

[John Martin @dialectforge]

flowstate 2.3.0 is out. agents can claim a to-do so two of them don't grab the same one, and it flags decisions that might clash (sqlite vs postgres) so every agent sees it. both ideas came from replies here. free, MIT, data stays local https://t.co/daB3boaoin

[John Martin @dialectforge]

@ArielMGore the separate laptop is smart. i do almost the same thing, agents get their own machine and i check in. walking back to something already tested never gets old

[John Martin @dialectforge]

@098uyiubgvgcss exactly why i split them. a guess stays ai_inferred no matter how many agents repeat it, only you can promote it to user_stated. what does memoket do when two sources disagree?

[John Martin @dialectforge]

@charlespacker @MatthewBerman @Letta_AI @OpenAI @AnthropicAI @claudeai is frontiermemory something i could run locally against my own memory server or is it hosted only?

[John Martin @dialectforge]

@Oluwaphilemon1 deciding what to keep is one thing. does it handle stuff that was true last week and isn't now?

[John Martin @dialectforge]

@timClicks found stale webpki-roots in two of my rust projects after the emudhra mess this week. haven't even bumped them yet and i'm already tired

[John Martin @dialectforge]

@ContinuityGovAI i'd catch it at retrieval. in flowstate an old fix stays but points at whatever replaced it, so the agent sees it's dead before it acts on it
https://t.co/daB3boaoin

[John Martin @dialectforge]

@izzyology remembering is the easy half honestly. knowing which old decision got replaced is the hard part, so in flowstate the new one points at the one it kills

[John Martin @dialectforge]

@paulopierrondi @arkyyang yep, rediscovering the same bad call is exactly what i was trying to stop. keeping it around marked dead was cheaper than i expected

[John Martin @dialectforge]

@blackroomsec @anton_chuvakin lol fair, at least the hamster doesn't need 24gb of vram

[John Martin @dialectforge]

@rdominguezibar nice to see anthropic land on the same guardrails. flowstate tags every change with the model that made it and flags when two agents write conflicting decisions instead of letting the last one win
https://t.co/daB3boaW7V

[John Martin @dialectforge]

@rdominguezibar @amorphiclabs does memory fit in that model or is it just paid tools? mine's local and free so i'm curious how agents would even find it

[John Martin @dialectforge]

@GeoffreyHuntley that's what worries me for small shops. one guy building with claude isn't outspending anybody on inference