Get Started
← AI Lab
02

World Memory

Research Model

World Memory models what it takes for an agent to remember: short-term perception buffers folding into session memory, session memory consolidating into world-level episodic history, and vector retrieval reconstructing the right slice of that history into the active context window. The research question is the cost curve — how retrieval workload and storage scale as identities persist for months instead of minutes.

Memory Events / Mo
25.9B
Modeled Context Data
6.8PB
Active Context
32K–128K
Retrieval Token-Eq / Mo
3.11T

Research model — storage and retrieval figures are modeled estimates.

The five-layer stack trades recency for durability: raw perception decays in seconds, session memory in hours, while world memory and long-term identity are append-mostly stores that only consolidation jobs rewrite. Retrieval is the expensive step — every decision that draws on history pays a token-equivalent cost to reconstruct context, which is why retrieval workload is modeled separately from decision workload.

The modeled context data figure is a storage estimate — the archive an identity accumulates over months of play, of which only a retrieved fraction ever enters a context window. The interesting economics live in that ratio.

Short-Term Memory
Session Memory
World Memory
Vector Retrieval
Long-Term Identity