Learning to Continually Learn via Meta-learning Agentic Memory Designs — part 5/5

Cleaned-text code points [62661, 65278). Bibliography omitted here; source.md includes it.

Dynamic Mode. We evaluate the learned best memory design alongside manual baselines on ALFWorld using the agentic system powered by GPT-5-mini/medium . The first 70 tasks from the valid_seen dataset are used to generate trajectories for the Memory Collection Phase, while tasks from the valid_unseen dataset are used in the Deployment Phase, creating a gap between collected memory and tasks to be solved. During the Deployment Phase, the agentic system uses general_retrieve() to access relevant memory for each new task. After completing a task, the interaction log is incorporated via general_update() , producing a dynamic memory that evolves as the agentic system sequentially solves tasks.

## Appendix C Result Details

### C.1 Learning Processes on Other Benchmarks

### C.2 Results of Greedy Exploration

We also conduct a greedy memory design sampling baseline on ALFWorld, where at each learning step we only select the historically best-performing memory design for further exploration (Appendix B.3 ). To ensure a fair comparison, this greedy strategy is evaluated under the same exploration budget, i.e., it explores the same total number of memory designs as our method. The resulting exploration tree is shown in Figure 14 .

ALMA produces learned memory designs that achieve better performance and transferability ( Table 2 ) compared to Greedy Search, highlighting the importance of open-ended exploration patterns to explore better designs.

### C.3 Example Results of Learned Memory Designs

The Best-learned memory design in MiniHack.

Example of the retrieved knowledge context during the Deployment Phase in MiniHack.

langley00

## Instructions for reporting errors

We are continuing to improve HTML versions of papers, and your feedback helps enhance accessibility and mobile support. To report errors in the HTML that will help us improve conversion and rendering, choose any of the methods listed below:

Click the "Report Issue" ( ) button, located in the page header.

Tip: You can select the relevant text first, to include it in your report.

Our team has already identified the following issues . We appreciate your time reviewing and reporting rendering errors we may not have found yet. Your efforts will help us improve the HTML versions for all readers, because disability should not be a barrier to accessing research. Thank you for your continued support in championing open access for all.

Have a free development cycle? Help support accessibility at arXiv! Our collaborators at LaTeXML maintain a list of packages that need conversion , and welcome developer contributions .