Deterministic deletion is predictable, and predictable systems are reconstructable. If a buffer is always cleared at a fixed interval, an observer with timing access learns the shape of what was there.
The primitive
Randomized deletion treats the pruning schedule itself as a source of entropy. Fragments of the working set are evicted on non-deterministic intervals drawn per-fragment, so no two runs of the same input produce the same residency pattern.
- Fragment the working set into units smaller than a semantic span.
- Draw an eviction offset per fragment from a bounded random distribution.
- Evict independently, so the surviving set at any instant is never a coherent slice.
- Enforce a hard ceiling: every fragment is gone before the window closes.
Why it matters
The property you get is not just deletion, it is non-reconstructability under partial observation. An adversary who samples memory mid-window recovers fragments that cannot be reassembled into meaning, because the surviving set is randomised across semantic boundaries rather than truncated along them.
This is the mechanism at the core of the ephemeral meeting intelligence patent, and it generalises well beyond meetings — any streaming inference pipeline that must not retain can adopt the same eviction discipline.