What's this about?
A memory scheme for LLM agents that performs read/write operations without spending context-window tokens on the bookkeeping — "zero-token" memory management (HN 93 pts, 12 comments).
What exactly is the token cost of memory?
Most agent memory works by stuffing remembered facts back into the prompt, which eats the context window and costs tokens on every turn. Zero-Mem aims to handle storing and retrieving memories outside that token budget, so remembering more doesn't shrink the space left for actual work.
Why does it matter?
Context is the scarce resource behind this week's long-horizon and long-context-degradation threads. If memory can be managed without taxing the context window, agents can hold more state over longer tasks — directly relevant to the day's long-horizon-harness papers.