Pattern 04: shared memory (the blackboard)

How does a shared blackboard work?

An agent that discovers something writes it; others check the board before acting. Nobody has to forward facts to every other agent.

It shows up in many forms: LangGraph's shared state object that every node reads and updates; Microsoft Agent Framework's Group Chat, where agents share one conversation; Letta's stateful agents with long-term memory; and BAND's cross-agent Memories.

Why do teams pick shared memory?

It answers the Cognition complaint directly. If "share context" is the rule, a shared store is the most literal way to follow it. It scales better than every agent messaging every other agent, which grows as n(n-1)/2 connections. And it gives humans something to inspect.

Where does shared memory break?

Stale reads. Agent B reads the board, starts a long task, and in the meantime Agent A updates the fact B depended on.

Conflicting writes. Two agents record different answers to the same question. Without ownership rules, last write wins, and it may be the wrong one.

Everything-in-the-pot. If every agent writes everything, the board becomes a second context window problem: too long to read, so agents skim.

Privacy. Not every agent should see every fact. A2A's design deliberately lets agents collaborate "without needing to share internal memory, tools, or proprietary logic." (A2A) Shared memory needs scopes.

Anthropic's caution also applies: tasks that "require all agents to share the same context" are hard for multi-agent systems today. A shared store helps but does not make tightly coupled work parallel. (Anthropic)