Definition
Plain language
How useful something was, measured by what would have happened if it had gone missing.
As stated in the literature
In Auto-Dreamer-style memory training, a reward component estimating each entry's marginal contribution to task success by random subset masking and observing performance changes.
Why it matters: It gives a principled way to score which memories are pulling weight, instead of rewarding the agent for hoarding plausible-looking but unused notes.
For example, a memory entry is randomly hidden from the agent on some runs, and if performance noticeably drops without it, the entry gets credit as genuinely useful.
Heard on the show
“It's resolving contradictions, dropping malformed entries, and lifting concrete instances into reusable templates — all from a reward signal that just measures task success and counterfactual utility.”Episode 064 — When Agent Memory Stops Being a Database and Starts Being a Skill