manage block registry

Current agent memory systems lack reliable evaluation benchmarks and systematic architectural analysis, leading to distorted performance assessments and ad hoc design choices. This work addresses this gap by proposing, for the first time, a taxonomy grounded in four distinct memory architectures, examined through empirical studies across multiple large language model backbones. The study systematically evaluates these architectures in terms of semantic utility, benchmark saturation, model dependency, and memory overhead, uncovering fundamental limitations that explain why real-world performance consistently falls short of theoretical expectations. These findings provide critical empirical evidence and actionable insights for designing scalable, evaluable memory systems in artificial agents.

This work addresses key challenges in large language model (LLM) agents, including inefficient context assembly, unstructured memory management, and insufficient access control during runtime. To overcome these limitations, the authors propose a compile-time-defined, permissioned memory model that introduces a registry of pure memory blocks annotated with ownership tags, enabling block-level operations with zero prompting overhead. Structured context assembly is achieved through five composable primitives—promote, gate, write, evict, and rollback—augmented by provenance labels, non-evictable author flags, and pattern-based eviction policies. Experimental results demonstrate substantial performance gains: block grouping improves task success rates by tens of percentage points; relevance-based gating reduces prompting costs by 67.8% while recovering 83% of success rates; and even small models with the proposed intervention outperform larger, unmodified counterparts.

AI agents lack standardized registration mechanisms for cross-domain discovery, authentication, and capability sharing across cloud, enterprise, and decentralized environments. Method: We systematically survey three prominent agent registration paradigms—MCP, A2A, and NANDA—and establish a four-dimensional evaluation framework assessing security, scalability, authentication, and maintainability. This enables the first horizontal comparison of centralized, decentralized, and cryptographically verifiable metadata models. Contribution/Results: We propose a unified registration architecture integrating structured metadata (mcp.json), Agent Cards, and verifiable credentials (AgentFacts), incorporating GitHub-based identity verification, well-known URI discovery, and distributed directory services. We rigorously delineate the operational boundaries of each approach across deployment contexts, thereby establishing design principles and standardization pathways to advance interoperability in the AI agent internet.

Current persistent large language model agents lack effective memory governance mechanisms, rendering them vulnerable to contradictory information, privacy leaks, and outdated “zombie memories.” This work proposes MemArchitect—a memory governance layer decoupled from model weights—that introduces, for the first time, a rule-based memory lifecycle management framework. By leveraging a policy-driven rule engine, MemArchitect enables explicit control over memory decay, conflict resolution, and privacy preservation. Experimental results demonstrate that memories governed by MemArchitect significantly outperform unmanaged baselines in agent tasks, underscoring the critical role of structured memory governance in enhancing the reliability and safety of autonomous systems.

Formal verification of operating system kernel virtual memory management (VMM) code remains challenging due to hardware interface complexity and difficulties in semantically modeling dynamic multi-address-space switching. This paper addresses these challenges by introducing a modal-logic-based abstraction of address spaces. Our method features: (1) a novel modal assertion ([r]P) to express truth relative to an address space (r); (2) a precise virtual *points-to* relation that faithfully models hardware page-table translation semantics; and (3) the first fully mechanized formal verification—within the Iris separation logic framework and Coq—supporting instruction sequences spanning multiple address spaces. We verify critical VMM operations including address-space switching and page-table updates. All semantic definitions and proofs are entirely mechanized in Coq, achieving significantly stronger verification guarantees than prior approaches.

This work addresses a critical security vulnerability in large language model (LLM) agents, where untrusted external inputs can be disguised as high-authority user history within long-term memory, leading to improper authorization of high-risk actions. To mitigate this, the authors propose PPMF, a lightweight memory middleware that preserves platform-maintained provenance metadata during memory compression and formally introduces the notion of a “provenance non-amplification bound” to prevent untrusted content from being “laundered” through memory integration. PPMF enables dynamic, runtime access control by combining provenance tracking, risk labeling, and structured memory representations based on action risk and memory authority. Experimental results demonstrate that under a fixed risk policy, PPMF completely blocks all unauthorized high-risk operations—reducing attack success rates to 0%—while preserving full functionality for benign and low-risk memory usage.

Current agent memory systems lack systematic evaluation of malicious content persistence, propagation impact, and remediation mechanisms. This work proposes MemSecBench, a benchmark that, for the first time, enables end-to-end security assessment of memory systems within a unified framework using a Write–Execute–Forget protocol in isolated environments. The benchmark spans seven lifecycle checkpoints and 24 agent configurations, supporting diverse memory backends and large language models. It introduces deterministic write validation, checkpoint-specific adjudicator models, and a programmatic gating mechanism. Experimental results reveal that 84.2% of configurations exhibit persistent malicious memories, with 50.3% completing full attack chains; among successful poisoning cases, 59.6% execute complete operation sequences, and 56.1% are amenable to selective remediation.