Short, sourced, job-shaped
A routing agent needs current territory rules, account ownership, recent intent, open exceptions, and the last operator correction. It does not need every conversation the company has ever had about segmentation.
// AGENT OPS · MACHINE-WORKER MANAGEMENT · June 6, 2026
Agent Ops Need Memory Budgets Before Machine Workers Hoard Context
Autonomous workers should remember enough to act well, not enough to become uninspectable.
Infinite context is not intelligence. It is unmanaged liability.
The lazy answer to agent performance is to give every machine worker more memory: more previous tasks, more customer history, more operator notes, more Slack context, more CRM state, more call transcripts, more exceptions, more everything.
That feels safe because the agent has fewer excuses. It is usually the opposite. When every decision can be influenced by a long, messy trail of stale facts, private notes, outdated policy, and accidental precedent, operators lose the ability to explain why the machine worker acted.
Memory budgets are the missing control. They define what an agent may retain, how long it may retain it, what must be summarized, what must be forgotten, and what requires a fresh source check before action.
Big, vague, sticky
An agent that carries old objections, informal promises, stale ICP notes, and one-off executive exceptions into every task starts enforcing ghosts. It looks smart while quietly turning history into policy.
The problem shows up as agent confidence with weak receipts.
A sales-assist worker recommends moving an account from expansion to churn-risk because it remembers an unhappy implementation thread from six months ago. Since then, the customer renewed, usage improved, and the CSM changed the health score. The agent is not hallucinating. It is worse: it is using real context with the wrong shelf life.
Now the operator has to debug memory, not output. Which fact mattered? Was it still valid? Was it allowed to affect this decision? Did the agent check the source of truth, or did it trust its own compressed history?
That review drag kills the whole point of machine labor. Agents should create leverage. They should not create forensic memory work.
Scope
What categories of memory can influence this job, and which systems are authoritative?
Shelf life
How long can a remembered fact survive before the agent must refresh it?
Weight
Which memories are hard constraints, weak signals, warnings, or irrelevant noise?
Visibility
What memory must appear in the receipt so operators can inspect the decision?
Deletion
What gets forgotten after the task, after correction, or after policy changes?
Memory budgets make machine workers easier to promote.
Operators do not need every agent to be brilliant. They need agents whose decisions can be traced, corrected, and safely widened. A memory budget turns context from a mystery blob into an operating surface.
Context hoarding
The agent sees a task, pulls every related record, compresses it into a persuasive answer, and claims confidence. The receipt says what it decided, not which memories were allowed to matter.
Budgeted recall
The agent lists the five memory classes it used, the source freshness, the excluded stale facts, and the operator correction that shaped the rule. Promotion becomes a control decision, not a vibe.
This is especially important in GTM systems, where old context decays fast. Buyer committees change. Product usage changes. Territories change. Pricing exceptions change. A machine worker with unrestricted memory can become extremely competent at enforcing yesterday.
The goal is not more memory. The goal is the right memory, with an expiration date and a receipt.
A practical memory budget for RevOps agents.
Operational memory
- Current system-of-record fields with timestamps.
- Active operator-approved rules and decision rights.
- Recent corrections tied to the same workflow.
- Open exceptions, unresolved disputes, and live stop conditions.
- Outcome signals from the agreed scoring window.
Legacy context
- Old chat summaries with no owner or source.
- One-off executive preferences not converted into policy.
- Customer sentiment older than the workflow shelf life.
- Prior agent decisions that were never reviewed.
- Any memory the worker cannot show in the receipt.
The budget does not make the agent dumb. It makes the agent honest about what shaped the decision. That honesty is what lets operators hand over more work without handing over the company’s institutional memory to a black box.
The memory review should be part of agent management.
When an agent makes a bad call, do not only correct the output. Correct the memory budget. Was the stale fact allowed? Was a fresh lookup missing? Did an old exception become hidden policy? Did the agent overweight a weak signal because it sounded specific?
Every correction should update one of three things: the allowed memory class, the shelf life, or the receipt requirement. Otherwise the same worker will repeat the same mistake with more confidence.
Concrete takeaway for this week
Pick one machine worker that uses customer, account, pipeline, or policy context and give it a written memory budget.
- Name the memory classes: define exactly what the worker may remember for this job.
- Set expiration rules: force refreshes for facts that decay, especially customer status and revenue ownership.
- Require source freshness: make the agent show where the deciding facts came from and when they were checked.
- Ban invisible precedent: old exceptions do not become rules unless an operator converts them into policy.
- Review memory after errors: fix the budget, not just the bad answer.
The blunt rule: if a machine worker cannot show which memories shaped the decision, it is not ready for more autonomy.