Shared Agent Memory¶
What It Is¶
Agents only collaborate if context survives between them. Shared memory is the substrate that carries it:
- Shared state / scratchpad / blackboard — a workspace object persisted across hand-offs, holding the brief, intermediate artifacts, and decisions. Without it, each agent starts blind.
- Short-term memory — the current session/run's working data.
- Long-term memory — knowledge persisted across runs: brand voice, ICP, and past-campaign outcomes.
- Hand-off protocol — explicit rules for what one agent passes to the next and when control transfers.
The Hand-off Is the Failure Point¶
The leading cause of multi-agent failure is not weak models but ambiguous hand-offs — state lost or misread between agents. Reliable orchestration therefore invests as much in the shared-state schema and hand-off contract as in the agents themselves.
How It Applies to Marketing Factory¶
Shared memory is what makes agent-orchestration actually work end-to-end: the campaign brief and accumulating artifacts live in shared state as work passes from content to design to landing-page agents. For the factory, the KB itself is long-term memory — brand voice, ICP, and especially the experiment logs that feed back as priors (closing the loop with the calibration-loop and experiment history). Getting the hand-off contract explicit is the cheapest reliability win available, and it is the same discipline whether the topology is hierarchical or swarm. It is the connective tissue of the marketing-factory.
Related Concepts¶
- agent-orchestration — shared memory is the substrate orchestration runs on
- workflow-vs-agent — simpler workflows need simpler shared state
- marketing-factory — the KB is the factory's long-term memory
- calibration-loop — logged outcomes persist as priors in long-term memory
Referenced from: marketing-agent-orchestration