How one team built shared memory for a multi-model AI agent fleet

A development team running a fleet of AI agents — each backed by a different language model — found that standard per-agent context windows caused repeated work, as no single agent retained knowledge discovered by another. The core problem was that passing more conversation history failed because a large context window on one model cannot be accessed by agents running on different models. Their solution was a model-agnostic shared memory system built on two layers: an append-only event log as the source of truth, and a derived index that agents actually query. Key design requirements included attribution of every memory entry to a specific agent and session, strict scoping so agents only access relevant context, and full auditability for regulated workflows. The team noted that most retrieval failures were actually scoping failures, and that narrowing the searchable candidate set per agent proved more effective than improving ranking algorithms alone.