How does CPython manage memory?

CPython uses reference counting as its primary mechanism: every object tracks how many references point at it, and it is freed immediately when that count hits zero. A generational cycle collector runs on top to catch reference cycles that counting alone cannot free. Memory freed by small objects usually returns to internal allocator pools rather than to the operating system.

Reference counting does most of the work. Every object carries a count, binding a name bumps it, letting a name go drops it, and at zero the object is destroyed right then. That determinism is why a file closes as soon as the last reference disappears in CPython, though I would not rely on it in code meant to run under other implementations. Counting cannot free cycles, so a parent holding a child that holds the parent back never reaches zero. That is what the generational collector handles, with three generations, young objects checked most often and survivors promoted. In a long running worker I debugged, memory climbed steadily and it turned out to be a cache holding objects with back references plus an exception traceback kept in a module level variable, which pinned entire frames. Tracemalloc snapshots taken ten minutes apart pointed straight at it. Worth adding that freeing objects often does not return pages to the operating system, so a flat high RSS after a spike is normal rather than a leak.