Design a Memory Index Loaded Every Session

PROJECT CONTEXT:
My agent's memory grew past what fits in context, but most of it is irrelevant to any given task. The index pattern from minimalist memory systems solves this: always load a small index, lazily load details when the task needs them.

GOAL:
Split memory into an always-loaded index and on-demand detail files, with the agent trained to follow index pointers.

REQUIREMENTS:
- A compact index file (hard token cap) always loaded at session start
- Detail files per topic, referenced from the index with one-line descriptions
- A memory_read tool for loading detail files on demand
- Index entries carry enough signal for relevance decisions without opening details
- Index maintenance: new topics added, stale pointers pruned

CONSTRAINTS:
- The index never exceeds its cap - details always live in detail files
- Every detail file must be reachable from the index or it does not exist

DEFINITION OF DONE:
- Session start loads only the index; context cost is small and constant
- Given a relevant task, the agent loads the right detail file unprompted
- Given an irrelevant task, no detail files are loaded
- The index accurately reflects the detail files after multiple sessions of updates

COMMON FAILURES TO AVOID:
- An 'index' that grew into the very memory dump it was meant to replace
- Detail files the agent never loads because index descriptions are too vague
- Orphaned detail files invisible to the index
- Loading every detail file every time, defeating the entire design