pagemem / KV memory
Eight conversations.
One pool.
A 512-token system prompt is shared by every chat. Preallocation gives each conversation its own copy of the maximum possible tail. Paging shares the prompt once and allocates the tail on demand. The ninth user is where the difference stops being a diagram.
The decision
With 504 pages in the pool, can you admit one more conversation without evicting someone?
receipt / 204 vs 504
Page wall / live allocator
PAGED + COW / 8 SESSIONS
shared prompt
preallocated padding
free page
session tail
The capacity read
—
—
—
pages returned vs preallocation
—
same pool / 200-token tails
—
external fragmentation
—
conversations admitted
Source receipts/receipt.json
Baseline 8 sessions / 512-token shared prompt
Pool 504 pages / 16-token input
Open receipt ↗