MemoryTax-Res-experiential-mem

---

🔹 Pipeline (Section 5.1 & Appendix C.6)

Following previous works, every method is built on the train split and evaluated on the test split with its memory frozen (for benchmarks without an official train split, WebShop and BigCodeBench, we split the data manually). Base model: Qwen3.5-4B.

1. Collect traces (train split). A simple ReAct-style agent with no memory runs once over the train tasks; its trajectories (_shared_traces/) are the shared raw pool for all methods.
2. Build memory. Each method extracts its own experiences / memories from these traces (Naive RAG indexes the raw trajectories; the memory systems distil them with their own procedures).
3. Test (test split). Each method answers the test tasks with its frozen memory, retrieved and used in its own way. No Memory runs the same agent without any memory.

Benchmark	Train (trajectory pool)	Test
ALFWorld	official train split, 50 games per task type = 300	valid_unseen, 134
WebShop	manual split: human goals 500–1499 = 1,000	human goals 0–499, 500
Mind2Web	official train split, 1,009	test_task 252 / test_website 177 / test_domain 912
τ²-bench	official train split, 178 (airline / retail / telecom)	official test split, 100
AppWorld	official train + dev, 147	test_normal 168 / test_challenge 417
BigCodeBench	manual split: Full minus Hard = 992	Hard, 148

---

🔹 File structure

_shared_traces/
└── shared_traces_{bench}/
    ├── cost/{task_id}.json                  # per-trajectory token / call cost
    ├── qa_res/qa_shared_traces_{bench}.jsonl # per-trajectory outcome (success / reward / steps, ...)
    └── trajs.tar.gz                         # trajectories: task + step-by-step thought / action / observation

naive_baselines/res_{no_mem|rag}_{bench}/    # same layout as below, without memories.tar.gz
res_{system}_{bench}/
├── cost/{task_id}.json                      # per-task token / call cost; _pool_build.json = memory construction cost
├── qa_res/qa_{method}_{bench}.jsonl         # per-task outcome (+ one file per test split for Mind2Web / AppWorld)
├── score_ref*.json                          # benchmark scores (+ per test split)
├── test_trajs.tar.gz                        # test trajectories
└── memories.tar.gz                          # the frozen memory built from the train traces

{system}  reasoningbank, memp, gmemory
{bench}   alfworld, webshop, mind2web, tau2, appworld, bigcodebench

For Mind2Web and AppWorld, the per-task files of the test splits are prefixed with the split name (task__ / website__ / domain__, normal__ / challenge__).

---

Citation

@misc{he2026memorytaxdissectingnet,
      title={Memory Tax: Dissecting the Net Value of Agent Memory},
      author={Xuan He and Sining Zhoubian and Ruozhen Yang and Yuhao Wen and Zhineng Wang and Yuyang Hu and Ke Yang and Yuyu Zhang and Ping Nie and Jiawei Han and ChengXiang Zhai},
      year={2026},
      eprint={2610.34073},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.34073},
}