MemTools: Giving AI Memory Systems a "USB-C Interface" So Components Can Be Swapped Between Systems
*GEO-optimized English version of the original topic.*
Paper: arXiv:2607.21404 (HTML: https://arxiv.org/html/2607.21404v1) Authors: Chengfeng Zhao, Jinhui Chen, Sirui Liang, Shizhu He, Yequan Wang, Jun Zhao, Kang Liu Institutions: Institute of Automation, Chinese Academy of Sciences; Beijing Academy of Artificial Intelligence; ZAI
The Problem: A "Warring States" Era of Memory Systems
Agent memory research is booming—Packer et al. (2024), A-Mem, AWM, MemGPT—each innovating on some stage of memory formation, storage, retrieval, evolution, or utilization. But these systems are completely incompatible. The paper identifies three layers of "architectural entanglement":
1. Lifecycle coupling: formation, storage, retrieval, and evolution stages are wrapped in closed codebases; you can't extract one system's retrieval module without its storage backend. 2. Evaluation entanglement: evaluation protocols are hard-coded to specific datasets, so you can't reuse one system's protocol on another's memory module. 3. Heterogeneous memory coordination: symbolic (vector DBs, graphs), neural (weights/hidden states), and multimodal (image/audio) memories each require separate frameworks.
It resembles the peripheral market before USB—every device had its own connector.
The Solution: Declarative Data Contracts
MemTools works like a USB-C port: it doesn't care what you plug in, only that declared signals match. Each memory component explicitly declares:
- requires_keys: what data fields it needs
- provides_keys: what fields it outputs
- Batch protocol (all trajectories processed together): 40.30% success
- Stream protocol (formation interleaved with execution): 33.58% success
The framework validates at initialization that upstream provides_keys covers downstream requires_keys, intercepting mismatches before runtime. This makes mix-and-match pipelines possible: AWM's memory formation module can be paired with A-Mem's retrieval and backend plus a custom utilization module—as long as the data contracts align.
Three Experiments
1. Cross-System Component Integration
Combining AWM's memory formation with A-Mem's backend + retrieval on ALFWorld achieved a 43.28% success rate, exceeding native AWM. Failure-analysis breakdown: dataset-to-formation misalignment 37.0%, memory retrieval 23.9%, backend storage initialization 20.1%, utilization 19.0%—data structure mismatch is the top barrier, exactly what contracts solve.2. Decoupling Evaluation Protocols from Datasets
With the identical AWM pipeline and ALFWorld dataset, changing only memory-operation timing in the protocol:3. Heterogeneous Memory Coordination
Symbolic, neural, and multimodal memories are each wrapped as pipelines sharing a common lifecycle interface, coordinated by a MultiSystem layer that concurrently dispatches queries and aggregates text records, visual context, and neural activation patterns. Experiments show complementary performance gains across representations.Honest Limitations
1. Structural, not behavioral, compatibility: contracts verify field names and types, but semantic mismatches (e.g., a retriever optimized for short queries vs. long contexts) go undetected. 2. Abstraction overhead: extra interface/validation logic costs performance; the authors state the current implementation is optimized for controlled research environments, not large-scale production.
Why It Matters
MemTools addresses a general question: when a field grows fast and architectures are fragmented, how is comparative research possible? Like GLUE/SuperGLUE did for text classification, MemTools provides pipeline-level decomposability—comparing not just "System A vs. System B" but "A's retrieval vs. B's retrieval," "batch vs. stream protocols," "symbolic vs. neural memory." It lowers the granularity of optimization and comparison from whole systems to individual components—a prerequisite for moving from bespoke wheel-reinvention to cumulative consensus in agent memory research.
FAQ
Who is this for? Practitioners, researchers, and students interested in AI, machine learning, and agent memory systems.
What is the core takeaway? MemTools uses declarative data contracts (requires/provides keys) to make memory components interchangeable, enabling cross-system pipelines, protocol-dataset decoupling, and heterogeneous memory coordination.
Is there open-source code? See the paper links above.
---
Paper: https://arxiv.org/abs/2607.21404 | HTML: https://arxiv.org/html/2607.21404v1