Skip to content

Token Budget

The share of the context window allocated to recalled memories, after the system prompt, tools, and conversation have taken theirs. Typically a few hundred tokens for something that may hold thousands of facts.

Why it matters in a memory layer

The budget is what turns retrieval into a selection problem. With room for six memories out of thirty-six, ranking stops being a nicety and becomes the whole game — and it is why measuring whether a memory earned its slot matters more than raising k.

Connections

Taught in: Context Is Not Memory

Used in: Getting Memories Into the Prompt · Summarization and Compaction · The Packing Problem · Why Forgetting Is a Feature