Skip to main content

Shared Memory

A common memory space that multiple agents can read from and write to, enabling coordination and information sharing without direct agent-to-agent communication. Agents post intermediate results, observations, and conclusions. Other agents read from the shared space to inform their own processing.

Also known as the blackboard pattern in classical AI.

Structure​

All agents have read and write access to a shared data structure. There is no central controller — agents independently decide when to read and what to contribute. The shared memory becomes the coordination mechanism.

Mechanism​

  • Agents write intermediate results, observations, and conclusions to the shared space
  • Writes can be structured (key-value) or unstructured (append-only log)
  • Optional namespacing to organize by topic, task, or agent
  • Conflict resolution: last-write-wins, merge, or versioned entries
  • Write access may be governed by policies (who can write what)

Key Characteristics​

  • Decoupled coordination — agents communicate through data, not direct messages
  • Asynchronous — agents read and write independently without blocking
  • Flexible — any agent can contribute any type of information
  • Consistency challenges — concurrent writes can create conflicts
  • Only useful in multi-agent systems — single agents don't need shared memory

When to Use​

  • Multiple agents need to collaborate on a shared task
  • Agents should be loosely coupled — no direct inter-agent communication
  • Intermediate results from one agent inform another agent's work
  • You need a coordination mechanism simpler than a central orchestrator
  • The task benefits from asynchronous, incremental contributions from multiple agents