EvoMemBench

EvoMemBench evaluates memory mechanisms in LLM-based agents across knowledge and execution tasks, both within an episode and across episodes. This repository provides the paper's evaluation subsets as typed Parquet tables for the Hugging Face Datasets library and automatic Croissant metadata export.

Use this version with datasets.load_dataset or the Hugging Face dataset viewer. For the benchmark's original JSON, JSONL, CSV, and YAML files, use EvoMemBench-original. Both versions contain the same evaluation subsets.

For the complete evaluation protocol of this dataset, please refer to the EvoMemBench GitHub repository

Load a configuration

Sign in with an account that has access to this private dataset:

pip install -U datasets huggingface_hub
hf auth login

from datasets import load_dataset

data = load_dataset(
    "littlewyy/EvoMemBench",
    "crossep_know",  # choose a configuration from the table below
    split="test",
    token=True,
)
print(len(data))  # 884
print(data.column_names)

Choose a configuration below. All nine configurations use the test split.

Configuration	Rows	Contents
inep_know	14	MemoryAgentBench contexts with 1,200 question/answer pairs in total
inep_exec	200	BFCL tasks joined with answers and four 50-task environment groups
crossep_tool	200	The same BFCL tasks, used by the cross-episode evaluation protocol
bfcl_tools	162	BFCL tool definitions from 12 libraries
crossep_know	884	CL-Bench tasks across 120 contexts, including complete messages and rubrics
crossep_know_difficulty	120	Context-level CL-Bench difficulty annotations
crossep_web_xbench	100	XBench DeepSearch tasks, preserving their original encoded text
crossep_web_webwalkerqa	170	WebWalkerQA questions, answers, root URLs, and navigation information
crossep_emb	200	ALFWorld test task indices; game assets are required separately

inep_exec and crossep_tool evaluate 200 unique BFCL tasks using the within-episode and cross-episode protocols, respectively. Tool definitions and difficulty annotations are supporting tables, separate from the evaluation tasks.

Working with the records

Most fields are ready to use as Python strings, lists, or dictionaries. Fields ending in _json hold JSON strings; decode them with json.loads when needed.

Data	Key fields and usage
InEp-Know	Each row has one context, a list of questions, and an aligned list of answers; each answer is a list of acceptable strings. metadata_json contains the complete context metadata.
BFCL tasks	question contains turns of {role, content} messages; ground_truth contains lists of function-call strings. Decode initial_config_json for the environment state. group and group_row_index identify the environment and its evaluation order.
BFCL tools	library, name, and description describe each tool. Decode parameters_json and, when present, response_json for its schemas.
CL-Bench	messages includes all conversation turns and rubrics lists the evaluation criteria. Group tasks by metadata.context_id and keep context_episode_index order. Use that context ID to join the difficulty table.
WebWalkerQA	question, answer, and root_url describe the task; info contains navigation and source details.
ALFWorld	item_id, task_type, and task_id identify each evaluation task. See the setup instructions below for the required game resources.

For example, read a BFCL task's initial environment state:

import json
from datasets import load_dataset

tasks = load_dataset("littlewyy/EvoMemBench", "inep_exec", split="test", token=True)
task = tasks[0]
initial_state = json.loads(task["initial_config_json"])

Keep question/answer pairs aligned in InEp-Know; QA IDs can repeat across contexts and should not be used to deduplicate them. XBench's id, prompt, answer, reference_steps, and canary columns retain their original string values, including encoded text and the canary marker.

Record identifiers and exact reconstruction

record_id is a stable row identifier. source_file and zero-based source_row_index locate its record in the original-format dataset; original_source_row_index also records an InEp-Know context's position before subset selection. Rows and nested lists preserve source order.

For exact BFCL reconstruction, source_fields, answer_fields, and definition_fields list the original keys, preserving the distinction between an absent optional field and an explicit null. A missing tool response is a null response_json; an explicit JSON null is the string "null".

For reproducible experiments, pass revision="<dataset commit>" to load_dataset, using a commit from Files and versions. MANIFEST.json contains the file inventory, counts, and checksums.

ALFWorld environment setup

The crossep_emb configuration contains 200 evaluation task indices. Obtain ALFWorld game resources separately and set ALFWORLD_DATA before running the environment. Download the accompanying configuration with:

from huggingface_hub import hf_hub_download

config_path = hf_hub_download(
    repo_id="littlewyy/EvoMemBench",
    filename="resources/alfworld/base_config.yaml",
    repo_type="dataset",
    token=True,
)

The benchmark's evaluation mapping uses task paths under json_2.1.1/valid_train. Follow that mapping when selecting games; the YAML's standard train/validation paths do not replace it. When using the benchmark's environment server, keep the code repository's training mapping file as well, because the server also loads it during setup.

Sources and usage

• MemoryAgentBench: the paper's selected retrieval and conflict-resolution contexts (14 contexts, 1,200 QA pairs).
• BFCL: 200 multi-turn tasks, answers, environment groups, and supporting tool definitions.
• CL-Bench: 884 tasks from 120 contexts with at least five episodes each, and their difficulty table.
• XBench DeepSearch: the 100-task evaluation CSV, preserving its encoded text and canary column.
• WebWalkerQA: the 170-task evaluation subset.
• ALFWorld: 200 evaluation indices and the environment configuration, excluding game assets.

Linked paper

EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective