Skip to content

MemTools: A Unified Research Framework for Interoperable Agent Memory

🕒 Published (v1): 2026-07-23 15:04 UTC · Source: Arxiv · link

Why this paper was selected

Unified interoperable agent memory framework; targets architectural fragmentation across memory lifecycle stages

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

MemTools is an interoperability research framework that decouples agent memory pipelines from deployment environments by standardizing the memory lifecycle through declarative data contracts, orthogonally separating evaluation protocols from benchmark datasets, and unifying symbolic, neural, and multimodal memory under a shared computational interface. It enables controlled isolation of individual memory design variables—previously impossible due to pervasive architectural coupling in existing systems. Empirical studies show cross-system hybrid pipelines can outperform native designs.

Problem

Existing agent memory systems (MemGPT, Mem0, MemOS, etc.) entangle three distinct concerns: (1) lifecycle stages (formation, storage, retrieval, evolution) are buried in closed codebases with no standardized inter-stage data contracts; (2) evaluation logic is bound to specific benchmark datasets, making it impossible to distinguish algorithmic performance from evaluation-setup artifacts; (3) heterogeneous memory representations (symbolic, neural, multimodal) lack a shared interface, preventing concurrent coordination. This architectural fragmentation prevents systematic, controlled comparison of memory design choices.

Method

MemTools introduces three interlocking design mechanisms:

  1. Standardized Lifecycle Interface: Each lifecycle stage (formation, retrieval, evolution, utilization, backend) exposes typed provides_keys/requires_keys declarative contracts. A matching engine automatically validates structural compatibility across stages and enumerates valid pipeline combinations without manual configuration, enabling cross-system component assembly.

  2. Unified Evaluation Paradigms: Benchmark datasets are reduced to pure Sample data providers; evaluation protocols (declarative offline, batch procedural, streaming online, cross-agent transfer, adversarial noise, etc.) are pluggable modules dictating execution timing independently of data. A single dataset can run under multiple protocols, isolating timing effects from memory algorithm effects.

  3. Heterogeneous Memory Management: Each memory modality (symbolic vector/graph DBs, neural KV-cache/activation steering/LoRA adapters, multimodal image/audio) is encapsulated in an independent pipeline conforming to the lifecycle interface. A MultiSystem coordination layer dispatches queries concurrently via synchronized parallel search, then applies retrieved components through a chained utilization pass to the base LLM.

Key Contributions

  • Declarative data contract system with automatic pipeline compatibility checking that intercepts structural mismatches across all lifecycle stages before execution.
  • Orthogonal decoupling of benchmark datasets from evaluation protocols, enabling configurable execution timing as an experimental variable.
  • MultiSystem coordination layer enabling concurrent operation of symbolic, neural, and multimodal memory pipelines under identical lifecycle interfaces.
  • Empirical demonstration that cross-system hybrid pipelines (e.g., AWM formation + A-Mem backend/retrieval) can exceed native architectures on ALFWorld (43.28 vs. 40.30 success rate).
  • Open-source framework with adapters for 10+ existing systems (MemGPT, Zep, Mem0, MemOS, A-Mem, etc.).

Results

  • Cross-system integration: AWM formation + A-Mem backend/retrieval hybrid achieves 43.28 success rate on ALFWorld vs. 40.30 for native AWM (Batch protocol). Incompatibility analysis across 4 lifecycle stages: formation–dataset (37.0%), retrieval (23.9%), backend (20.1%), utilization (19.0%).
  • Protocol isolation: Identical AWM pipeline on ALFWorld scores 40.30 (Batch) vs. 33.58 (Stream), attributing a 6.72-point gap entirely to memory formation timing rather than algorithmic differences.
  • Heterogeneous coordination (symbolic + multimodal): A-Mem + Omni-SimpleMem coordinated yields F1 = 35.3 on Mem-Gallery vs. 30.1 (A-Mem alone) and 31.2 (Omni-SimpleMem alone).
  • Heterogeneous coordination (symbolic + neural): AWM + NPM coordinated yields 51.5 success rate on ALFWorld vs. 48.5 (AWM alone) and 32.1 (NPM alone).

Limitations

  • The matching engine validates only structural field compatibility; subtle behavioral misalignments between cross-system components are not detected.
  • Abstraction layers introduce computational overhead compared to tightly coupled native systems.
  • Current implementation targets controlled research environments and may encounter performance bottlenecks when coordinating massive heterogeneous databases over very long task horizons.
  • Pairwise modality coordination is tested; full three-way (symbolic + neural + multimodal) coordination is not empirically evaluated.

Relevance to Agentic AI / LLM Agents

MemTools directly addresses a core infrastructure problem in agentic AI research: the inability to rigorously compare or recombine memory mechanisms across systems. By treating memory lifecycle stages, evaluation protocols, and memory representations as independently configurable variables, it enables the kind of controlled ablation studies that are standard in other ML subfields but rare in agent memory. The finding that evaluation timing alone (Batch vs. Stream) shifts ALFWorld success by ~6.7 points has direct implications for how the community should interpret published memory benchmarks. The framework's support for symbolic/neural/multimodal coordination is particularly relevant as agents move toward richer, multi-sensory environments requiring heterogeneous long-term memory.