Summary

Position paper framing multi-agent memory as a computer architecture problem. Identifies two critical unsolved gaps: (1) cache sharing across agents without consistency guarantees, (2) structured memory access control at the protocol level.

Source: arXiv 2603.10062 — Multi-Agent Memory from a Computer Architecture Perspective: Visions and Challenges Ahead
Zhongming Yu et al. Published 2026-03-09.

Key Insights

• Three-layer memory hierarchy: I/O layer (tool outputs) → cache layer (working memory/context) → memory layer (persistent store)
• Gap 1: No standard cache-coherency protocol for multi-agent memory — different agents may have stale or conflicting cached facts
• Gap 2: Access control is application-level hack, not protocol primitive — no reliable way to enforce read/write boundaries across agent trust boundaries

Applicability to Zeph

This is a vision paper — no algorithm to implement directly. Value is as an architectural framing for future design decisions.

Mapping to Zeph's stack:

• I/O layer ≈ tool outputs in tool-output/ overflow dir
• Cache layer ≈ context_manager working window + ContextBuilder
• Memory layer ≈ SQLite key_facts/episodes + Qdrant embeddings

Design implications:

1. When designing research(memory): collaborative memory with multi-agent access control (CollabMem) #1864 (collaborative memory), model access control as protocol-level metadata, not application-level filtering
2. Subagent memory reads should go through a cache coherency check — is root agent's memory still valid since subagent last read it?
3. A2A agent discovery should advertise memory capabilities (read/write/scope) as part of agent card

Use case: Reference this framing when designing multi-agent orchestration memory sharing in #1840 (AdaptOrch) and #1864 (CollabMem).

Action

No direct implementation. Use as reference architecture when:

• Designing subagent memory isolation (see research(memory): collaborative memory with multi-agent access control (CollabMem) #1864)
• Adding A2A memory capability advertisement to agent card
• Planning cache invalidation strategy for shared memory pools