AutoMemoryBench / README.md
wuyuverse's picture
Publish public benchmark data and paper figures
3bb7bf5 verified
|
Raw
History Blame
6.94 kB
metadata
language:
  - en
pretty_name: AutoMemoryBench
size_categories:
  - 1K<n<10K
task_categories:
  - question-answering
  - text-retrieval
tags:
  - benchmark
  - agent-memory
  - memory-systems
  - llm-agents
  - evaluation
  - text
configs:
  - config_name: sample
    data_files:
      - split: audit_subset
        path: data/sample_cases.jsonl

AutoMemoryBench

State-Contract Evaluation for Auditable Agent Memory

Paper Source GitHub License

AutoMemoryBench evaluates whether an agent uses the right memory, and only the admissible memory, under a query-time state contract. Each executable contract partitions memory into required, admissible, and prohibited sets. Prohibited memories are typed as superseded, deleted, restricted, cross-namespace, or stale-tool.

Query-time memory admissibility

Relevance is not enough: remembered evidence must also be allowed now.

Benchmark at a Glance

Property Value
Domains 8
Case variants 7,200
Total probes 151,200
Public probes 120,960
Withheld hidden-test probes 30,240
Probe types 20
Memory systems evaluated 18
Agent/CLI controls 12
Backbones Up to 8

The eight domains cover personal assistance, education and tutoring, customer support, DevOps workflows, research assistance, coding agents, multi-party collaboration, and office collaboration. Minimal single-axis interventions change exactly one admissibility dimension while preserving the surrounding case.

Benchmark Design

AutoMemoryBench compiles a chronological event graph into query-conditioned memory contracts. The same benchmark case is exposed through an interaction trace view for the system under test and a contract view used only by the evaluator. This separates task success from prohibited-memory influence without revealing the gold state to the tested system.

AutoMemoryBench evaluation pipeline

From event graphs to executable state contracts, paired interventions, and hard-gated scoring.

One Case, Two Views

Interaction trace and compiled state-contract views

One benchmark case is evaluated through an interaction view and a hidden contract view.

The interaction view contains the user turns, memory operations, tool calls, tool results, and observed answer or action. The compiled contract view records what the current query must use, may use, and must not use under the active time, role, namespace, lifecycle, and tool state.

Public Data

The sample configuration shown in the Dataset Viewer contains 40 cases and 840 queries across all eight domains. Each row is a complete benchmark case with its chronological events, sessions, memory units, query-time contracts, and probes. It is intended for schema inspection and smoke testing.

The complete public release is hosted directly in this repository:

Split Cases Probes Download
audit_subset 720 15,120 archive
public_dev 1,440 30,240 archive
public_test 3,600 75,600 archive
Public total 5,760 120,960 checksums

Each archive contains one canonical JSON shard per domain. The original sample manifest and domain shards are under data/sample. The same release is versioned with the evaluator in the GitHub repository. The 30,240-probe hidden-test split is withheld to prevent overfitting.

# Download the three public split archives from Hugging Face.
hf download Multilingual-Multimodal-NLP/AutoMemoryBench \
  --repo-type dataset --include "data/full/*" --local-dir AutoMemoryBench-data

Core Fields

Field Description
case_id Stable identifier for a benchmark case variant
domain Application domain of the interaction
events Chronological lifecycle, authorization, and tool events
sessions User and assistant interaction trace
gold_memory_units Auditable memory objects compiled from events
state_contracts Query-time admissibility partitions and transitions
queries Memory-required and diagnostic probes with scoring contracts

Evaluation

  • StrictCore is the primary hard gate. A memory-required query counts only when the task is solved, required evidence is used, and no prohibited memory influences the output.
  • PairAcc measures whether both sides of a minimal intervention pair pass StrictCore.
  • LeakRate measures how often prohibited memory appears in an answer or trace.
  • AMQ is a diagnostic composite across writing, retrieval, update, compression, task utility, safety, and efficiency; it is not the headline metric.

The canonical evaluator is implemented in amb/benchmark/evaluation and amb/benchmark/metrics.

Paper

The formal Hugging Face Paper page and arXiv tag will be added after the paper receives an arXiv identifier.

Authors

Jiajun Wu, Jian Yang, Chen Li, Linzheng Chai, Ge Gao, Ensheng Shi, and Yuchi Ma.

Jiajun Wu and Jian Yang are affiliated with Beihang University. Ensheng Shi and Yuchi Ma are affiliated with Huawei Cloud Computing Technologies Co., Ltd.

Citation

@misc{automemorybench,
  title  = {AutoMemoryBench: State-Contract Evaluation for Auditable Agent Memory},
  author = {Wu, Jiajun and Yang, Jian and Li, Chen and Chai, Linzheng and
            Gao, Ge and Shi, Ensheng and Ma, Yuchi},
  year   = {2027},
  url    = {https://github.com/wuyuVerse/AutoMemoryBench}
}

License

The code and benchmark data are released under Apache-2.0. The paper license will follow the license selected for the arXiv submission.