If you have to re-brief your agent every day, it’s not a colleague – it’s a chore. Memory changes that.
Priyanka Mudgal
Principal, AI Engineering
Workday
The way work gets done in HR and finance is rapidly changing. AI agents now source candidates, reconcile invoices, flag payroll anomalies and co-ordinate multi-step processes.
Gartner predicts 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% just a year ago. By 2028, at least 15% of day-to-day work decisions are expected to be made autonomously by agentic AI.
That second number is worth sitting with. Agents won’t just assist decision-making – they’ll make a number of meaningful decisions. This means agents are increasingly going to operate as part of the workforce and need to be managed the way an organisation manages any other member of that workforce: with context, accountability and a clear record of what they know and why.
By 2028, at least 15% of day-to-day work decisions are expected to be made autonomously by agentic AI.
Guide
An agent without memory is a new hire every single morning.
It can reason impressively in the moment, but it can’t recall that the same supplier invoice was flagged twice last quarter, that a hiring manager prefers evidence-backed candidate summaries, or that the last time it approved an exception like this one, a human reviewer reversed the decision. Every interaction starts from zero – and in HR and finance, where the stakes of a wrong decision are high, zero context is not good enough.
Consider a finance agent monitoring expense submissions. Without memory, it evaluates every report in isolation, blind to the fact that this is the third time this month a particular vendor’s invoices have arrived just under the approval threshold.
A human controller would notice the pattern immediately. An agent that starts fresh each time never will, no matter how sophisticated its reasoning is in any single instance. Digital labour only earns trust when it can be managed with the same discipline organisations already apply to people, finance and operations – and that discipline starts with what an agent is allowed to remember, for how long and under whose oversight.
This is why it is crucial to treat memory as core infrastructure for agentic AI, not a nice-to-have feature layered on top.
Agents only earn trust when they can be managed with the same discipline organisations already apply to people, finance and operations.
Great decision makers don’t just think well – they remember well. They carry forward lessons from past outcomes, patterns from similar situations, and context about the people and processes involved. The same principle holds for a digital workforce: memory is what turns a capable agent into a dependable one.
Independent research backs this up. Agents built with dedicated, structured memory consistently reason more accurately over long, multi-session work than agents that simply rely on a larger context window – and they do it faster and at a fraction of the computational cost.
That efficiency gain is the practical case for memory as infrastructure, not just a research curiosity. Replaying entire histories gets slower and more expensive, and important details still get lost in the noise. Structured memory is what makes agentic AI something you can actually run at enterprise scale.
For a digital workforce, there are several key benefits:
1. Memory gives agents continuity across a process, not just a conversation. A hiring cycle, a financial close or a benefits escalation unfolds over days or weeks, across many touchpoints – not one chat session. An agent that remembers can reopen a paused requisition, recall why a candidate slate was narrowed last week, and carry forward the preferences a hiring team already set, instead of re-asking questions the business already answered.
2. Memory lets agents learn from what actually happened, not just from the prompt in front of them. Did the forecast adjustment hold up? Did the escalation resolve the issue? An agent that stores outcomes alongside its decisions can recognise a situation that resembles one that went wrong before – and adjust the same way a good employee builds judgment over time.
3. Memory lets agents work the way teams actually work. Approval preferences, communication styles, jurisdiction-specific requirements – this is the operational texture that makes a recommendation feel like it came from someone who knows the business, not a generic assistant.
4. Memory reduces the tax employees pay for supervising AI. When an agent has to be re-briefed on context it should already know, the person working with it ends up doing double duty – completing the task and re-teaching the agent at the same time. Memory is what lets that supervision shrink over time instead of staying constant no matter how long the agent has been deployed.
None of this requires an agent to be more “intelligent” in the way that term usually gets used, but more continuity is needed. The unlock comes from context, not raw capability – which is exactly why memory deserves attention as its own discipline.
An agent without memory is a new hire every single morning.
Here’s the part that’s easy to get wrong: the goal isn’t for an agent to remember everything. It’s for an agent to remember the right things.
Recent AI research illustrates why: most memory-augmented systems are surprisingly bad at catching contradictions in what they’ve stored, quietly acting on outdated or conflicting information rather than flagging it.
Left unchecked, a bad memory doesn’t just sit there – it quietly shapes every decision that follows, the same way one uncorrected error ripples downstream. Most memory systems today aren’t built to catch that “error compounding effect” before it compounds.
Picture a benefits agent that once approved an exception for an employee’s unusual medical circumstance. If the agent stores that exception as a general rule rather than a one-time judgment call, it may quietly start applying it to every similar-looking case that follows – cases that were never actually similar, just superficially close. No one decided the policy should change. It changed anyway, one unreviewed memory at a time.
An agent with no memory makes inconsistent decisions, but an agent with bad memory makes consistently wrong decisions, confidently. That is what makes ungoverned memory riskier than no memory at all.
The question we should be asking is: will this fact change a future decision? If the answer is no, it doesn’t belong in AI memory.
Treating memory like a well-curated knowledge base, rather than a junk drawer, is what keeps an agent’s judgment sharp instead of cluttered.
An agent with no memory makes inconsistent decisions, but an agent with bad memory makes consistently wrong decisions.
The stakes of getting this right are already visible in the market. Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Unmanaged agent memory sits squarely inside all three.
Four guiding principles matter here:
1. Memory must be auditable. When an agent’s recommendation is shaped by something it learned last month, leaders need to be able to trace that influence – what was stored, when and why it shaped the next action.
2. Memory must be correctable. Policies change, org structures change, people change roles. It has to be as easy to update or retire a stored fact as it was to create one, so an agent doesn’t confidently act on yesterday’s truth.
3. Memory must respect boundaries. Compensation data, one-off anomalies and anything outside an agent’s mandate should be filtered out at the door. A recruiter’s general hiring preferences are not the same thing as a decision made for one confidential backfill, and neither should leak across users, roles or agents.
4. Memory must be scoped to the right owner. Not every fact an agent learns belongs to the organisation broadly. A hiring manager’s personal working style and a company-wide compliance requirement both live in memory, but they carry very different weight and should never be treated as equally durable or equally shareable across teams.
None of this should slow down agent adoption, but managing memory shouldn’t be an afterthought. Organisations that build these habits early are the ones that can scale agentic AI past a single pilot use case without governance debt catching up to them. The ones that skip it tend to discover the gap only after an agent has already acted on something it should never have retained.
Intelligence gets an agent started. Memory, which is managed with the same discipline you’d apply to any member of your workforce, is what makes it wise.
Guide
Discover how CIOs across ANZ are turning AI investment into measurable business value with trust built in.