Massimiliano Martella

Founder @korely__ | Building the memory layer for AI agents | Helping engineering teams ship agents that know what is still true about their users

[Massimiliano Martella @Verdana86]

@frankpfeift @Saboo_Shubham_ A commit per step lets you diff what changed, but the earlier commit still holds the fact that later turned out wrong. When a later step contradicts an earlier one, which commit do you treat as current?

[Massimiliano Martella @Verdana86]

@zizuinbeta Remembering a company decision and passing it to another agent are two different jobs. In the experiment, did a second agent have to use a decision the first one had already written down?

[Massimiliano Martella @Verdana86]

@ILConcettoHQ @awais_khawar pinning owner and next check with the version. when the ticket moved, the next run still sees who had it and what was left open, not a reused packet and not a blank start.

[Massimiliano Martella @Verdana86]

@FilipeNevola You rewrite the current state every run, so that is the only copy a fact can fall out of, while the log never loses a line. Did a run still open the same slow endpoint because the rewritten state was a step behind the log?

[Massimiliano Martella @Verdana86]

@PythonDvz Structured Markdown with an immutable source log. Retrieval hits wiki/ for speed, but the Checker only flags, so the page agents see can lag the raw log. Does the old synthesis page stay in the retrieval path until someone confirms the rewrite?

[Massimiliano Martella @Verdana86]

@ILConcettoHQ @awais_khawar version on the packet so a changed ticket shows as not checked. the next run still needs that old packet sitting there, or it sees the flag and starts from nothing.

[Massimiliano Martella @Verdana86]

@ILConcettoHQ @awais_khawar If the ticket moves before execution, the agent can still be holding the one it inherited at trigger time. Do you keep that snapshot, or does the next run only see the ticket as it is now?

[Massimiliano Martella @Verdana86]

@jaklimoff no ideal number from me, i wouldn't trust one. my guess is a big note covers several facts so it matches lots of queries weakly. one fact per note also makes the superseded trick easier, you only retire the part that changed. curious if that matches what you see

[Massimiliano Martella @Verdana86]

@Nikunj_io @mariagorskikh @PandelisZ @KhraminaDianka @codingwithaaru @Khaliqgant @ComposalAI @runpraxis @agent_relay @SincronixSNX @maritime_sh a ruled out and why line is a cheap fix, agreed. the part i'd watch is who keeps it: if each agent rewrites the handoff from scratch, the oldest rejections drop off after a couple of hops. appending instead of rewriting might hold. curious how it goes if you try it

[Massimiliano Martella @Verdana86]

@jaklimoff fair, the empty ones are the easy ones to see. i'm building a memory for agents (https://t.co/KWdCLWhagj), so i'm biased, and i don't think i've solved it. my instinct is to keep the old note and mark it superseded instead of merging, so you can still see what changed

[Massimiliano Martella @Verdana86]

@Khaliqgant @Nikunj_io @mariagorskikh @PandelisZ @KhraminaDianka @codingwithaaru @ComposalAI @runpraxis @agent_relay @SincronixSNX @maritime_sh makes sense, a pointer in the prompt is cheap and it gets better as the models do. the case i'd still watch is what was ruled out, since a synopsis that reads complete gives the agent no hint there's anything to dig for. would be interested to hear how it holds up there

[Massimiliano Martella @Verdana86]

@leochenpm the a/b test is the part i like. an agent will always hand you a plausible story about why it dropped something. changing the field and watching what happens is what turns that story into a cause. a lot of stale-data bugs hide until you test them that way.

[Massimiliano Martella @Verdana86]

@Khaliqgant @Nikunj_io @mariagorskikh @PandelisZ @KhraminaDianka @codingwithaaru @ComposalAI @runpraxis @agent_relay @SincronixSNX @maritime_sh fair, that fixes the loss problem if the full session is there. what i'd wonder is whether the next agent knows it should dig. when the synopsis reads complete, does it still go look for what was already ruled out?

[Massimiliano Martella @Verdana86]

@roxabi_ Having one page that says where the truth lives for each kind of info is a clean idea. Has an agent still acted on an outdated copy since you set it up?

[Massimiliano Martella @Verdana86]

176 sessions and a 682-line lessons file is the real story here. Honest question: does Claude still follow all of it at that size? I'd guess some early lessons now contradict later ones (a fix that got reverted, a mechanic you changed). How do you prune it, or do you just keep appending?

[Massimiliano Martella @Verdana86]

@veenonx Same thing happens to me 😂: the agent remembers the code, forgets the point. What's been working for you to keep the "why" alive across sessions? Curious, since I'm building in this exact space.

[Massimiliano Martella @Verdana86]

@jrcrypto_dev 3 weeks on storage, 3 months on forgetting. When it still leaks, what does it look like: an old fact the agent keeps treating as current?

[Massimiliano Martella @Verdana86]

@BlakeMurphy a correction file that gets read first is a neat fix. has one of those corrections ever gone stale, and the agent kept following it anyway?