AuthorsZhongming Yu, Naicheng Yu, Hejia Zhang et al.
TL;DR
Multi-Agent Memory Architecture reframes multi-agent context as a three-layer memory hierarchy plus cache and access protocols to ground future consistency models.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Multi-Agent Memory Architecture argues that multi-agent systems face a bottleneck that mirrors computer architecture: memory hierarchy, bandwidth, and consistency constraints.
Without structured shared or distributed memory, collaborative LLM agents risk stale reads, overwrites, and incoherent context, undermining long-horizon reasoning and safety.
HOW IT WORKS
Multi-Agent Memory Architecture introduces a three-layer hierarchy: Agent IO Layer, Agent Cache Layer, and Agent Memory Layer, plus Agent Cache Sharing and Agent Memory Access Protocol.
Think of Multi-Agent Memory Architecture like CPU I O, cache, and DRAM, where agent context flows through fast working memory into slower but larger persistent stores.
This KEY_MECHANISM lets Multi-Agent Memory Architecture reason about multi-agent context as data movement and consistency, something a plain context window cannot express or control.
DIAGRAM
This diagram contrasts how Multi-Agent Memory Architecture conceptualizes shared and distributed memory for coordinating multiple LLM agents.
DIAGRAM
This diagram shows how Multi-Agent Memory Architecture layers Agent Context IO, Agent Cache Sharing, and Agent Memory Access Protocol over the memory hierarchy.
PROCESS
Agent IO Layer
Multi-Agent Memory Architecture first routes user inputs through Agent IO Layer so audio, text, images, and network calls become structured context.
Agent Cache Layer
Multi-Agent Memory Architecture stores compressed context, recent trajectories, and tool calls in Agent Cache Layer for low-latency multi-hop reasoning.
Agent Memory Layer
Multi-Agent Memory Architecture persists full dialogue history and external knowledge database entries in Agent Memory Layer for long-term retrieval.
Agent Cache Sharing
Multi-Agent Memory Architecture uses Agent Cache Sharing and Agent Memory Access Protocol so other agents can safely transform and reuse cached artifacts.
KEY CONTRIBUTIONS
Agent memory hierarchy
Multi-Agent Memory Architecture defines Agent IO Layer, Agent Cache Layer, and Agent Memory Layer to mirror computer memory hierarchies for LLM agents.
Protocol extension for multi-agent scenarios
Multi-Agent Memory Architecture proposes Agent Cache Sharing and Agent Memory Access Protocol to move beyond raw connectivity like MCP and JSON RPC.
Multi-agent consistency agenda
Multi-Agent Memory Architecture identifies multi-agent memory consistency as the next frontier and calls for formal consistency models and verification frameworks.
RESULTS
Memory layers
3 layers
Agent IO Layer, Agent Cache Layer, Agent Memory Layer
Protocol gaps
2 protocols
Agent Cache Sharing and Agent Memory Access Protocol
Memory paradigms
2 types
Shared memory and distributed memory architectures
Consistency models
3 tiers
Architecture memory, agent memory, multi-agent memory comparison
Multi-Agent Memory Architecture is a position paper without benchmarks, so the key numbers describe the proposed hierarchy, protocol gaps, and consistency comparison. These counts show how Multi-Agent Memory Architecture structures the design space rather than reporting accuracy gains.
BENCHMARK
Multi-Agent Memory Architecture is a position paper without benchmarks, so the key numbers describe the proposed hierarchy, protocol gaps, and consistency comparison. These counts show how Multi-Agent Memory Architecture structures the design space rather than reporting accuracy gains.
BENCHMARK
Relative emphasis across memory layers and protocol gaps in Multi-Agent Memory Architecture.
KEY INSIGHT
Multi-Agent Memory Architecture argues that multi-agent systems are already hitting a memory wall, even without explicit token-length failures or benchmark collapses.
This is counterintuitive because many builders still treat prompts as static strings, underestimating how bandwidth, caching, and coherence silently limit multi-agent performance.
WHY IT MATTERS
Multi-Agent Memory Architecture gives researchers a shared vocabulary to design agent memory like computer architects design caches and consistency models.
Builders can now reason about Agent Cache Layer, Agent Memory Layer, Agent Cache Sharing, and Agent Memory Access Protocol instead of ad hoc prompt hacks.
Peilin Feng, Suorong Yang, Soujanya Poria
arXiv 2026 · 2026
Σ-Mem maintains historical competence evidence matrices M_p, a peer relationship matrix G, a memory readout r_{p,t}, and residual steering δ_{p,t} to guide multi-agent trust decisions. On the mixed counterfactual benchmark, Σ-Mem raises Qwen3-0.6B accuracy at CF@90 from 46.22% with the base model to 71.10% with Σ w/ G.
Xiaoyang Li, Yiqi Wang et al.
arXiv 2026 · 2026
Correlated Promotion Benchmark (CPB) combines CPB-Static, CPB-Live, a gold admission rule, lineage collapse, and a governance rule to stress-test epistemic admission in shared agent memory. On CPB-Live, the governance rule keeps damage shares between 0.112 and 0.152 and false adoption between 0.06 and 0.09, while majority vote and LLM judges often match share-all’s false adoption.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Answers use this explainer on Memory Papers.
Checking…