MemTools: A USB-C Interface for AI Memory Systems — Interchangeable Components via Declarative Data Contracts

Forum topic · ✨步子哥 · 2026-08-03

Summary

MemTools (arXiv:2607.21404), developed by Chengfeng Zhao's team at the Institute of Automation, Chinese Academy of Sciences, is a framework that standardizes how AI agent memory systems connect, analogous to a USB-C port for hardware. Current agent memory research — including A-Mem, AWM, and MemGPT — suffers from three layers of architectural entanglement: lifecycle stages locked inside closed codebases, evaluation protocols hard-coded to specific datasets, and incompatible heterogeneous memory representations. MemTools introduces declarative data contracts: each component declares requires_keys and provides_keys, and the framework validates compatibility automatically. Experiments on ALFWorld show that a mixed pipeline combining AWM's memory formation with A-Mem's backend and retrieval achieves 43.28% success, beating native AWM. Isolating protocol timing reveals batch processing (40.30%) outperforms streaming (33.58%). The framework also coordinates symbolic, neural, and multimodal memories via a MultiSystem layer. Limitations include no behavioral compatibility checks and abstraction overhead.

What is MemTools?

MemTools is a framework that gives AI agent memory systems a standardized interface so components from different systems can be swapped and combined — the paper's authors compare it to a USB-C port for peripherals. The paper (arXiv:2607.21404, published July 23, 2026) comes from Chengfeng Zhao's team at the Institute of Automation, Chinese Academy of Sciences, with collaborators from BAAI and the Zhongguancun Institute of Artificial Intelligence.

The problem: a fragmented landscape

Agent memory research is active — from Packer et al. (2024) to A-Mem, AWM, and MemGPT — but systems are mutually incompatible across three dimensions:

1. Lifecycle coupling: memory formation, storage, retrieval, evolution, and usage are bundled into closed codebases; you cannot extract a single module. 2. Evaluation entanglement: evaluation protocols are hard-coded to specific datasets. 3. Heterogeneous memory silos: symbolic (vector DBs, graphs), neural (weights/hidden states), and multimodal memories cannot be coordinated.

The solution: declarative data contracts

Each component declares:

requires_keys — the data fields it needs
provides_keys — the fields it outputs

The framework validates at initialization that upstream provides_keys cover downstream requires_keys, enabling mix-and-match pipelines and failing fast on mismatches.

Key experimental findings

Cross-system integration: AWM's memory formation module combined with A-Mem's backend + retrieval achieved 43.28% success on ALFWorld, beating native AWM. Failure attribution: dataset-to-formation alignment 37.0%, memory retrieval 23.9%, backend storage init 20.1%, usage stage 19.0%.
Protocol decoupling: identical pipeline and data, only timing changed — Batch protocol scored 40.30% vs Stream protocol 33.58% (a 6.72-point gap previously invisible due to protocol-dataset coupling).
Heterogeneous coordination: symbolic, neural, and multimodal memories run as parallel pipelines under a MultiSystem coordinator, yielding complementary performance gains.

Honest limitations

Data contracts verify structural compatibility only, not behavioral/semantic compatibility (e.g., short-query vs long-context retrieval modules).
The abstraction layer adds computational overhead; the implementation is optimized for controlled research settings, not large-scale production.

Why it matters

MemTools provides pipeline-level decomposability: instead of only comparing "System A vs System B," researchers can compare individual components, protocol timings, and memory representations. This mirrors how GLUE/SuperGLUE enabled comparability in NLP. It signals the memory systems field moving from a fragmentation phase toward cumulative, comparable research.

FAQ

Who is this for? Practitioners, researchers, and students in AI, machine learning, and deep learning, especially those building agent memory systems.

Where can I read the paper?

Paper: https://arxiv.org/abs/2607.21404
HTML: https://arxiv.org/html/2607.21404v1

Authors: Chengfeng Zhao, Jinhui Chen, Sirui Liang, Shizhu He, Yequan Wang, Jun Zhao, Kang Liu (Institute of Automation, CAS; BAAI; Zhongguancun Institute of Artificial Intelligence).