Campaign Memory

Campaign Memory

Paste a campaign brief or a results update and get a short campaign status back, while an agent keeps a memory graph of the campaigns, audiences, channels, claims, proof, assets and results it has heard about. The graph grows with every update, so the agent knows which claim is approved, which is paused and who owns each asset. It is a small, automatic version of an ontology: a map of a campaign drawn by a model and stored in your own database.

What this workflow does

1. brief is a text input with the brief or update. The default text is invented.
2. CampaignMemory is an Agent node with persistent memory switched on. Its instruction asks for seven kinds of thing (Campaign, Audience, Channel, Claim, Proof, Asset, Result), for the status of each claim (approved or paused) and for every link as a full sentence, such as "Sofia Lindqvist owns the Spring Cold-Chain Webinar".
3. Reply returns the campaign update.

After each run a second model call reads the text and the reply and merges the entities and links into the agent's graph. The graph is stored in two PostgreSQL tables, agent_memory_nodes and agent_memory_edges.

Setup

1. Import the template into Heym: download the JSON from this page and drag it onto the Workflows tab, or paste it.
2. On CampaignMemory, pick your credential and a model.
3. Run it with the sample brief, then replace the text and run it again with each update.
4. Start a message with QUESTION: to ask what the graph knows. The agent answers from memory only and says what is missing.
5. Click the brain icon on the Agent node to read and edit the graph.

What to expect

We ran this template on Heym 0.0.148 on 7 November 2026 with claude-haiku-5-5, using three sets of invented campaign text and one question. With the short vocabulary in the instruction, three runs ended at 14 to 16 entities and 15 to 19 links. The stored rows, the SQL queries and what the runs showed are in Auto Ontology: A Guide With Templates to Try. Its Sales Call Memory twin does the same for call notes.

FAQ

What does the agent remember? Entities such as people, campaigns, assets and claims, and the links between them. Heym stores them in two PostgreSQL tables that belong to the Agent node, and you can read, edit and delete them from the brain icon on the node.

Does the memory cost extra model calls? Yes. After each successful run a second model call updates the graph. In our twelve test runs of the two templates it used 1.5 to 2.7 times the tokens of the agent's own calls.

What happens when a fact changes? Some links are replaced and some are kept. In six campaign runs the old owner of an asset stayed linked beside the new owner every time, and a pause was recorded in five of the six. Read the graph once a week and delete what is wrong.

Is the sample text real? No. Every name, company and figure in the sample text is invented.