Replaying a real session to find the memory cost we never counted
Long sessions kept getting reset after our memory fixes. We nearly capped the context, until a local replay of a real session pointed at tokenization.
EngineeringRead articleHow we think about agents, business and engineering.
Long sessions kept getting reset after our memory fixes. We nearly capped the context, until a local replay of a real session pointed at tokenization.
EngineeringRead articleThe docs say 128 MB, yet our sessions crossed it without always being killed. We deployed a throwaway probe to find out: the reset line fell between 254 and 300 MiB, what matters is how much data is alive at once, and recorded peaks are always a lower bound.
EngineeringA user said our agent kept stopping halfway through. Our logs showed no cause of death; the platform metrics showed memory errors, and heap snapshots showed three sources. This is how we tracked it down, the traps we hit, and the numbers after the fix.
Engineering