... very manual, I am testing the behaviour first before I build anything real. I am still not sure about 2 things. If people actually want to be confronted with their own old decisions, and how much should happen automatically vs asking me to confirm every time. Would love to hear what you think, especially if you tried something like this and it did not work. Comment from Aravind Venkatesan(@aravind_venkatesan1): Creating files for each context is good for smaller ... ... rules can be static skills and instructions and the decisions made for case by case basis can be in the tool such as mem0. Comment from Uttam Kumar(@uttam_kumar35): The what changed? prompt is probably more valuable than perfect

... also remembers everything wrong. One stale or slightly-off fact gets written once, and then every agent inherits it with full confidence which is harder to catch than an obviously dumb answer, because the memory is doing the reassuring. Perfectrecall of unverified context isn't an edge, it's faster confabulation. So the bottleneck for me isn't the model or the volume of context it's trust in the context. Feeding it is cheap. Knowing which remembered thing ... ... write down. Comment from Omar Ghandour(@omarships): @mystoryland This is the sharpest version of the point and I think you're completely right. "Remembers everything" was the provocation, not the actual claim, because you nailed the failure mode:

... fintech startup and we've just raise $4m So we rebuilt our context system, to help Pretty Prompt give better results. (Best part, the user owns that context, and can use it across any LLM) Launching today Instructions vs Memory. Instructions Explicit rules that always apply. Examples: No em-dashes Keep answers under 512 characters Use bullet points Memory Background context that only appears when relevant. Examples: My audience is non-technical I'm launching on Product Hunt and like ... ... temporary, and what should act like a rule. One thing I d be curious about: do you let users inspect or correct the classification after the fact? I ve found that trust in AI memory often depends less on perfect