Pre-registration: When Does Learning Memory Pay? (MemBench two-regime holdout freeze)
This is a pre-registration / analysis-plan freeze for a controlled criteria study of learned memory managers for LLM agents (working title: When Does Learning Memory Pay?). It is not a new SOTA memory method. Confirmatory Table-B uses tau2-bench mock, holdout H=8, T in {5,10} (memory-necessary vs memory-optional), Qwen3-4B-Instruct, five seeds {44-48}, and a Bonferroni family of four contrasts. Phase-1 1-seed API cells are exploratory only. T in {20,40} is stretch pending a filler protocol.
Authors
- M. Zhang
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-08-18
- DOI
- https://doi.org/10.5281/zenodo.21991694
- Primary Topic
- Memory Processes and Influences
- Type
- preprint