A preprint submitted to arXiv on 12 November 2026 introduces MAPE, a framework for evolving complete AI agent memory programs. Xinheng Ding and colleagues report experiments on conversational-memory benchmarks that examine both what an agent retains and how it prepares retrieved evidence for an answer.
What MAPE changes in agent memory design
The authors’ code repository describes memory programs with three stages: building a memory from conversation history, retrieving evidence for a query and composing that evidence into answer context. A language model proposes program revisions, while a diagnosis agent examines execution traces to guide subsequent changes.
Two patterns recur in the authors’ analysis. Keeping original conversation material available alongside extracted memories can recover details that extraction omits. Their Query-Routed Scaffolding module organizes retrieved material according to the question’s intent before answering. The authors report accuracy improvements when they transplant that module into five existing memory systems.
This separates evidence selection from evidence preparation, a useful distinction for researchers studying retrieval-augmented generation. The result invites investigation of failures after retrieval as well as failures to retrieve the right material.
What the benchmark evidence covers
The MAPE paper evaluates LoCoMo and LongMemEval-S using separate search and evaluation partitions. Evolution uses the search partition; the evaluation partition subsequently supports final program selection and performance reporting. These are the authors’ experimental findings from a preprint.
| Benchmark | Scope described by its creators |
|---|---|
| LoCoMo | Generated long conversations refined by human annotators; tasks include question answering, event summarization and multimodal dialogue. |
| LongMemEval | Chat-assistant memory abilities including information extraction, reasoning across sessions and time, knowledge updates and abstention. |
The authors also compare accuracy-only evolution with an objective that considers serving cost. They report that including cost changes the allocation of computation, moving work from memory construction toward selective processing at query time. That finding concerns the tested configurations rather than an established effect across agent applications.
What is available for research teams
The repository includes the evolution loop, baseline systems, evaluation components and an evolved program case study. Its project-owned code uses the Apache License 2.0; bundled upstream implementations retain their own licences. Benchmark data are downloaded separately and retain their original release terms.
For universities and research offices assessing AI research outputs, the contribution is a testable account of memory design choices. In CASRAI’s assessment, the distinction between retained evidence and its composition offers a useful experimental question for institutional pilots. Performance on conversational benchmarks alone does not establish suitability for research administration workflows.
Sources
- Ding and colleagues, Discovering Effective Agent Memory Designs through Program Evolution — arXiv submission record. Checked 2026-11-15.
- MAPE preprint, full text. Checked 2026-11-15.
- MAPE authors’ code repository and licence information. Checked 2026-11-15.
- LoCoMo: Evaluating Very Long-Term Conversational Memory of LLM Agents. Checked 2026-11-15.
- Wu and colleagues, LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory. Checked 2026-11-15.
