Core concepts
Context assembly
How the context window is built for the way a model reads a prompt.
The context assembler builds optimized context windows for LLM consumption. It respects how transformers actually read a prompt, and only sends what changed.
U-curve ordering
Places high-salience memories at the start and end of the context window, where LLM attention is strongest. Lower priority items go in the middle. This primacy/recency layout maintains high compliance from the model on what matters most.
Token budget
Respects the configured max_tokens limit. Memories are ranked by salience and packed into the budget, ensuring no overflow. Purpose-built serialization formats (compact pipe-delimited) fit more memories per budget than raw JSON.
Delta serving
Tracks what was sent in previous turns. On subsequent calls, only new or changed memories are included, reducing memory retrieval tokens by roughly 90% across multi-turn conversations. Early turns carry a small delta-header overhead, but by turn 10+ savings exceed 85% per turn.
See Token usage and Benchmarks for the measured numbers.