Designing MemoryService for Multi-Agent Systems: Policy or Storage Layer?

Fun-Following-1723 · reddit · 2026-08-27

A junior developer is designing a MemoryService for a multi-agent system (Supervisor + specialized agents) that sits in front of providers like Mem0 or TencentDB. The post discusses architectural boundaries: should the application own memory policies (promotion, deduplication, scope mapping) while providers act as storage/retrieval layers? It raises concerns about conflicting logic if the application explicitly approves a memory but the provider's internal logic rejects it. The author seeks advice on whether to treat providers as dumb storage or if an adapter layer is necessary to map user/agent/shared scopes.

Original post →

More from coding & agent

coding & agent channel →