01 / evidence
Positive results and failure boundaries, together.
State improves the main Hermes search-query task. A separated target-frequency prior explains a larger share of the gain, while heldout tools and TAU-bench expose where it stops transferring.
Inspect every headline number →trace / poolresetstatebest
GLM-5.1 · all0.32670.36550.5522support proxy
Kimi · all0.33400.37890.4559prior + residual
MRR@10 · BGE-M3 · closed candidate vocabulary
02 / method
One retrieval score. Three auditable signals.
01
Prefix
Keep only actions and responses observed before the target.
02
Separate
Model causal state and train-only target frequency as different signals.
03
Stress
Report candidate coverage, tool-heldout losses, and clustered uncertainty.
03 / boundary
This is evidence, not a memory product claim.
URL retrieval is a negative control. TAU train-only coverage averages 11.92%. Support gating can lose on cold tools. The docs keep these limits adjacent to every gain.