Research Shows The Best AI Memory Isn’t Bigger. It’s More Selective.

In this blog we discuss:

• How AI Decides What’s Worth Remembering 

• Workday’s Methodology Yields Higher Quality and Faster Results 

• Explicit Facts Are the Clearest to Parse 

• 3 Reasons Agent Memory Matters for Enterprise AI 

• Memory Is a Design Decision, Not a Side Effect

As AI agents take on more tasks across more workflows, what they remember from past conversations—and what they forget—is key. An agent that can’t filter stores the wrong information until it's bloated, slow, and quietly seeded with errors. But an agent that filters too aggressively may discard the important details. The challenge for organizations is knowing what's worth keeping. 

Workday AI Research, our new team committed to making AI more trustworthy and useful, addressed this challenge in a recent study. They discovered a new method that made AI far more selective about what to keep—delivering 12% higher precision and a roughly 8% gain in overall memory quality. 

How AI Decides What’s Worth Remembering 

Most existing enterprise memory systems sit at two extremes. Simple rule-based approaches are fast but imprecise—they can't reliably stop made-up or unreliable information from being stored. Fully AI-driven approaches are more capable but slow, expensive, and hard to audit. In other words, organizations can't easily see why they kept what they kept.

Our researchers explored a different approach. For the purposes of the study, they looked at 1,500 agent memories, and asked AI to answer five plain-language questions before deciding what to keep and what to filter out:

• Is it useful? Will this likely matter in a future conversation?

• Is it trustworthy? Is it actually supported by what was said—or possibly something AI invented?

• Is it new? Or already stored?

• Is it fresh? How recently did it come up?

• What kind of information is it? Is it a durable fact, like a preference or an identity detail, or a passing state?

What makes this approach practical is that only the first question—usefulness—needs an expensive AI model. The other four are handled by fast, transparent rules. The final answers combine into a simple weighted score, so a developer can look at any memory  and see exactly why it was kept or dropped.

This method also handles conflicting information deliberately. When new information contradicts something already stored, it doesn't pile both up—it keeps the higher-value version and merges them, so the memory stays accurate instead of accumulating contradictions.

In the study, Workday’s research method retained 97% of the memories that mattered and posted the highest decision-quality score of any system tested—while running more efficiently. By comparison, the leading LLM-native alternative retained every relevant memory, but also admitted more irrelevant candidates, which resulted in lower precision.

How different approaches compare on the same benchmark, scored on how well they keep useful memories while filtering out the junk.

Workday’s Methodology Yields Higher Quality and Faster Results 

Workday’s new method delivered 12% higher precision on keep-or-drop decisions while sacrificing almost none of the memories that mattered. That translates to about 8% improvement in overall quality.

Workday’s methodology was also 31% faster than the leading system (roughly 2.6 seconds per candidate versus 3.8). The difference came down to ‘AI calls’—how often each method consulted the AI model: the rival made several per decision, while our method made just one. 

In fact, its four rule-based checks took less than 65 milliseconds combined; nearly all of the 2.6-second decision time came from that single AI call. This hybrid design saves that slower, more expensive step of querying the AI model when more judgment is needed.

Explicit Facts Are the Clearest to Parse 

The most revealing finding came from testing how much each of the five questions contributed to AI memory decision-quality. The single most important factor was   information type—whether something was a durable fact worth keeping or a passing state safe to forget. Removing that one question caused the largest drop in quality (roughly 18%), while the other four added smaller, complementary refinements.

The results of Workday’s study also varied by conversation type. AI did better with personal conversations, where people stated facts explicitly, like preferences, identity details, or family relationships. It struggled more with professional conversations because important context is often assumed, scattered across systems, or never said aloud.

This gap points to a broader enterprise lesson: an AI agent makes better decisions about what to remember when it has a clear, connected foundation of relevant information to draw on. The stronger the underlying data and context, the better an agent can identify what matters.

3 Reasons Agent Memory Matters for Enterprise AI

The research reveals three benefits of using a framework like the one tested by our researchers to improve agent memory:

1. Quality beats quantity. A useful agent memory isn't the one that stores the most information, it's the one that remembers the right things. Being selective at the moment information is stored keeps the system fast and its memory clean.

2. Lower risk from bad information. Before storing a memory, the system checked whether it was supported by evidence in the conversation, reducing the risk of unsupported information entering long-term memory.

3. Transparency you can audit. Every keep-or-drop decision is an inspectable score rather than a black-box judgment. Teams can see why a memory was kept or dropped, and adjust the criteria to fit their own context. This makes debugging, auditing, and accountability possible in a way opaque systems can’t match.

Memory Is a Design Decision, Not a Side Effect

As AI agents take on more tasks and roles, it's tempting to treat their memory as something that just accumulates in the background. However, the research argues that organizations need to take a more strategic approach. What an agent remembers—and what it forgets—should be a deliberate, inspectable choice. Get that right, and reliability and trust follow. 

A note on the research: A-MAC was evaluated on LoCoMo, a public benchmark of roughly 30 long conversations (personal-assistant, technical-support, and collaboration dialogues) that together yielded about 1,500 candidate memories with human-labelled decisions about which should be retained. It was not tested on Workday products, customer data, or workplace records; the enterprise parallels in this piece are illustrative. The quality figures are balance scores that weigh useful memories kept against irrelevant ones filtered out—they are not accuracy percentages.

Looking for more insights on how to make AI trustworthy, useful, and efficient for the enterprise? Visit workday.com/ai-research for a complete list of Workday AI research papers.