Free-Form, Structured, or Hybrid? Memory Form Determines What Lifelong Agents Can Reuse

External memory lets lifelong agents reuse experience without weightupdates, but after a trajectory is reflected into reusable experience it muststill be written in some form — a free-form note, typed structured facets,both, or nothing — and this write-time choice is usually hard-coded. Ina paired continual-learning study on ALFWorld (two separately sampledgrids, 16,200 trajectories, exact McNemar and paired-bootstrap statistics),we find that form, not memory presence, determines value. Free-formnotes are statistically null in every cell; raw trajectory replay, run in onegrid, significantly hurts on the seen split; and every significant gain comesfrom a method that writes structured facets, with the gains that replicateacross grids concentrated out of distribution. Structure has a predictablecost: always-structured memory flips previously solvable games to failuresprecisely in a pre-specified over-transfer cell, and an environment-tunedrule selector is significantly harmful on seen transfer. A per-trajectory LLMselector retains the benefit of structure without environment-specific tuning— statistically tied with always-structured memory in seven of eight head-to-head comparisons, and never significantly worse than no memory. It cut thepredicted over-transfer harm by 75% in one grid but not in the other, whichwe report strictly as a single-sample observation. A cross-model replicationunder a variant protocol sharpens the conclusion: for a stronger actionmodel, every fixed structure-bearing form turns significantly harmful onthe seen split, while the LLM selector posts that grid’s only significant gain.Memory form is thus a consequential, selectable design axis for lifelongagents — and the right form is capability-dependent, which favors selectionover hard-coding.