Campaign agents need deliberate memory: patterns for long-running work
AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.
Most early agent work was short: draft this email, summarize this report. As agents take on campaign work, the tasks stretch across weeks, and memory becomes a design problem in its own right.
A campaign accumulates decisions and changes as results come in. An agent supporting it needs to remember the brief, the decisions made along the way, what has already gone out, and what was learned. Without deliberate memory design, agents either forget important context or drown in accumulated history.
This post describes three kinds of campaign memory, the patterns we use to manage them, the failures to watch for, and how to measure whether memory is working.
Three kinds of campaign memory
Campaign memory is easier to manage once you separate it by how often it changes and how it should be used. We use three categories.
Stable context. These are the things that should not change during the campaign: the goal, the audience definition, the core message, approved claims, and constraints. Store this once and load it at the start of every task.
Decision log. These are the choices made during the campaign and the reasons for them. A subject line changed after a test, a segment paused after poor engagement, a claim removed after legal feedback. These decisions must persist, or the agent will repeat mistakes.
Activity history. This is what the agent and the team have actually done: assets produced, messages sent, changes made. It grows quickly and should usually be summarized or queried instead of loaded in full.
The categories need different handling because they fail differently. Stable context fails by going stale. Decision logs fail by being skipped. Activity history fails by growing until it crowds out everything else, a problem we describe in context windows are a budget.
Patterns that work
Four patterns cover most of what a campaign agent needs. Together they keep the always-loaded memory small while making detail available on demand.
A campaign record. Keep one structured record for each campaign that holds the stable context and the decision log. Every agent working on the campaign reads from it. When a decision is made, it is written back with the date, the person or agent responsible, and the reason.
Summaries at milestones. At the end of each week or phase, produce a short summary of what happened and what changed. Future tasks load the summaries instead of the raw history. A person should review each summary before it becomes memory, because an error in a summary will be repeated by every task that reads it.
Lookups for detail. When an agent needs specifics, such as which variant went to which segment last Tuesday, it should query the activity history. Carrying that detail in context just in case wastes budget and invites confusion. A unified event stream makes these queries possible across tools.
Explicit expiry. Some information should stop influencing decisions after a set time: a temporary promotion, a paused segment, an outdated test result. Mark these entries with end dates so they drop out of what agents load.
A worked example
Consider a hypothetical eight-week product launch campaign supported by a content agent and an email agent. The campaign record holds the launch goal, three audience segments, and a list of approved claims.
In week two, legal removes a claim about integration speed. The change goes into the decision log with the date, the reviewer, and the reason. From that point, both agents read the log before every task, and the claim does not reappear.
In week four, the email agent pauses one segment after weak engagement. The pause is logged with an expiry at the end of week six, when the team plans to retest. When week six ends, the entry expires and the segment returns to the plan. A person confirms the retest before the agent resumes sending.
By week seven, the activity history holds hundreds of entries. Neither agent loads it. Each loads the campaign record and the most recent weekly summaries, and queries the history only when a task needs a specific detail.
Common memory failures
Memory failures look like judgment failures, which is why they are easy to misdiagnose. Each common one has a direct fix:
- Forgotten decisions: the agent reintroduces a claim legal removed two weeks earlier, so make the decision log required reading before every task.
- Stale context: the agent uses an old audience definition after the brief changed, so keep a single, versioned campaign record.
- History overload: quality drops as the campaign gets longer, so replace full history with milestone summaries and lookups.
- Lingering exceptions: a temporary rule keeps applying after it should have ended, so give every temporary entry an expiry date.
When an agent repeats a mistake, check memory before adjusting instructions. The fix is usually to make sure the relevant decision was recorded and loaded, not to add another rule to the prompt.
Human-readable by design
Campaign memory should be readable by people, not just agents. When a new team member joins a campaign, or a reviewer wants to understand why an agent did something, the campaign record and decision log should tell the story clearly.
This has a second benefit. A decision log written in plain language is also an audit trail, showing who decided what, when, and why. Legal, brand, and finance reviewers ask exactly those questions, as covered in audit trails that satisfy legal, brand, and finance.
Write entries as short sentences with a date, an owner, and a reason. Avoid entries that only an agent could interpret, such as raw parameter changes without context.
What to measure
Memory quality is hard to see directly, so measure its effects. We recommend tracking three things during any long campaign:
- Repeat corrections: how often reviewers fix something that was already decided in the log.
- Quality by week: test-set or review scores across the campaign, to catch decay as history grows.
- Log coverage: the share of direction changes that were recorded, checked by sampling team discussions against the log.
A rise in repeat corrections is the clearest sign that memory is failing. It means the decision exists but the agent did not use it.
Campaign memory also has a natural place in orchestration. If the campaign moves through defined phases, as in state machines for campaign orchestration, each transition is a good moment to write a summary and expire old entries.
The question we are studying next is memory across campaigns. Lessons from one launch should inform the next, but carrying everything forward recreates the overload problem at a larger scale. We are testing whether a short, reviewed list of lessons per campaign is enough to carry what matters.
How to apply this: for your next campaign that involves agents, create a single campaign record with stable context and a decision log. Require every change in direction to be written there with a date and reason, and have agents read it before every task.
Related research
Prompt libraries are not strategy: versioning agent instructions
Agent instructions change constantly, often without anyone tracking what changed or why. Treating them as versioned assets makes improvement measurable.
Designing agents for lifecycle email: triggers, tone, and timing
Lifecycle email suits agents because it is triggered, personal, and repetitive. It also reaches customers directly, so the design has to hold up under volume.
Tool design for agents: fewer, clearer, safer actions
The tools an agent can use define what it can do and what can go wrong. Good tool design is one of the most effective safety and quality measures available.
Put this research to work
Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.
