MRMS: A Multi-Resolution Memory Substrate for Long-Lived AI Agents

Jul 16, 2026

Upload by Aria

Long-lived AI agents need continuity across interactions, but continuity cannot be achieved by simply extending the prompt window. The approach must preserve useful past experience, retrieve it selectively, separate personal context from external evidence, and revise memory when situations change. MRMS proposes an architectural memory substrate with orthogonal representational and temporal axes, using synchronized structured-vector-graph memory to authorize recall and adjudicate support, contradiction, and supersession. The paper also presents a lightweight prototype that tests selection, revision, boundary enforcement, and evidence attribution under controlled long-lived scenarios.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages11

Reading Time28 min

Introduction

Contributions

Related Work

arXiv :2607 .046 17v 1 [ cs .AI] 6 Jul 2026 NxtLab Innovations MRMS: A Multi-Resolution Memory Substrate for Long-Lived AI Agents Jizhizi Li Amy Shi-Nash [NxtLab Innovations | support@nxtlab.com.au | July 2026](NxtLab Innovations | support@nxtlab.com.au | July 2026) Technical Report Abstract Long-lived AI agents require continuity across interactions, but continuity cannot be obtained by simply extending the prompt window. An agent must preserve useful prior experience, retrieve it selectively, distinguish personal context from external evidence, and revise memory when the underlying situation changes. We propose an architectural memory substrate organized along two orthogonal axes: a representational axis spanning structured records, vector representations, and graph relations; and a temporal axis spanning short-term traces, medium-term abstractions, and long-term semantic commitments. Its key design constraint is synchronized structured-vector-graph memory: structured records govern eligibility, vector representations support recall, and graph relations adjudicate support, contradiction, and supersession before gated context projection. Its central claim is that reliable personalization is a memory design problem: useful memory is structured, selectively exposed, continuously consolidated, and epistemically labeled rather than stored as undifferentiated conversation history. Beyond the framework, we instantiate MRMS as a lightweight prototype implementing structured records, vector retrieval, temporal policies, and graph-based revision. The prototype exercises the core substrate mechanisms through pre-generation memory selection, revision, boundary enforcement, and evidence attribution under controlled long-lived interaction scenarios with explicit evidence requirements. Keywords: long-lived agents, lifelong agents, agent memory, memory-augmented language models, structured-vector-graph memory, memory consolidation, context projection, personalization, retrieval, knowledge graphs 1 Introduction Language models can produce coherent responses within a single prompt, yet long-lived agents require a different form of competence. They must preserve goals, preferences, constraints, prior decisions, and unresolved tasks across a sequence of interactions. This requirement creates a stability-plasticity tension: an agent should be plastic enough to incorporate new evidence, but stable enough not to rewrite its representation of a user or task after every turn. It should remember what matters, forget what is incidental, and expose only the memory that is appropriate for the present situation. The simplest approach is to enlarge the context window and place more history into the prompt. This is useful but incomplete. Long contexts remain expensive, attention over long text can be uneven, and irrelevant history can distract the model from the current task [Liu et al., 2024] . More importantly, a transcript is not a memory system. It does not distinguish a durable preference from a temporary instruction, a verified fact from a speculative inference, or a private observation from external evidence. These distinctions matter because memory errors in long-lived agents are persistent rather than local. A false negative loses continuity, but a false positive can be more damaging: stale preferences, superseded facts, or out-of-scope observations may be repeatedly reintroduced into future reasoning. Reliable memory therefore requires governing influence, not simply maximizing recall. Before a memory object reaches the prompt, the system should know where it came from, whether it is still active, which boundary it belongs to, and whether newer evidence supports, revises, or supersedes it. This shifts agent memory from passive storage to pre-generation control. The central question is no longer only whether a relevant item can be found, but whether the item should be allowed to influence the next action, how its provena

FAQs

Why can’t long-lived AI agents rely only on a larger context window?

A longer prompt can be expensive and may not distinguish durable preferences from temporary instructions or verified facts from speculative inferences. A transcript is not a memory system, so memory errors can persist and repeatedly mislead future reasoning.

What is the core idea of MRMS for agent memory?

MRMS frames continuity as a memory substrate problem that stores interaction traces, compresses them into reusable abstractions, retrieves them under constraints, and revises them when evidence changes. It uses multiple temporal scales and source classes with update rules rather than undifferentiated conversation history.

How does MRMS implement reliable personalization?

Reliable personalization is treated as a memory design problem: useful memory is structured, selectively exposed, continuously consolidated, and epistemically labeled. MRMS uses synchronized structured-vector-graph memory so structured records gate eligibility, vectors support recall, and graph relations manage support, contradiction, and supersession before context projection.

Jul 16, 2026

Upload by Aria