the myth that python "handles memory for you" is why your agents OOM at 4 hours of uptime

ran 24 multi-agents in parallel last month, burning 10x the tokens of a single session for ZERO usable output
the real problem wasnt the tokens though, it was the memory nobody was watching
python uses reference counting plus a cyclic garbage collector. sounds fine until you load numpy arrays through C-extensions that dont decrement refs properly. those objects NEVER get collected. they just sit there, growing, silent
every 100 tokens of context your long-running agent processes, thats another tensor allocation that might not release. multiply that by 24 concurrent sessions and youre leaking 400MB/hr on a good day
> just add more RAM
yeah thats $30k/mo in compute to compensate for something tracemalloc would have caught in 10 minutes.