AI agents are learning on the job — just not for your whole team

Every correction a worker makes to an AI agent vanishes the moment a colleague opens the same tool. That's not a bug — it's a design gap that explains why productivity gains remain elusive.

Written by Lena Armitage · Bureau Tech · Jun 6, 2026

The correction that disappears

Here is the problem in one sentence: when a worker refines an AI agent — better prompts, sharper feedback, more precise context — that refinement evaporates the moment a colleague opens the same tool.

The next person starts from zero. The agent has no memory of what the first person taught it.

This isn't a fringe edge case. It's the default behavior of most enterprise AI agents today, and it compounds badly in multi-agent workflows, where the expectation is that agents share context across users and tasks. Without a shared memory layer, every team member trains a different version of the same agent. Those versions never sync.

Why the productivity numbers don't add up

Asana's own research puts a number on the gap: 75% of knowledge workers use AI on the job, but only 5% of companies report measurable productivity gains. That's a striking disparity, and Asana Chief Product Officer Arnab Bose has a specific diagnosis.

"Model providers are getting really, really good at improving reasoning and retry loops, but what they're not good at is bringing the enterprise work context in a way that human beings can reason about for shared memory," Bose told VentureBeat.

Asana's response is its Agentic Work Management platform, which routes corrections through a shared context graph — so if one team member improves an agent's behavior, that improvement applies to everyone. "You don't have to have every human member of the team become an expert at prompt engineering or context engineering," Bose said.

That's a meaningful claim, though it's worth noting it comes from a vendor with a product to sell. Independent validation of how well the shared context graph performs at scale isn't yet available.

A structural problem, not just a prompt problem

Sriharsha Chintalapani, co-founder and CTO of data governance firm Collate, frames the issue as one of consistency. "Agents are sensitive to the quality of their prompts," he told VentureBeat. "Someone with a strong understanding of the task will generally get more accurate results than someone less experienced." The agent remembers corrections from that one user — and only that user.

Chintalapani argues organizations need to stop treating shared memory as a prompt engineering problem and start building systems that propagate context across every conversation. Neej Gore, chief data officer at Zeta Global, describes the goal as a "living memory that compounds intelligence across the enterprise."

The technical challenge is real. The models powering agents are stateless by design — they don't retain information between sessions. Memory has to be implemented as a dedicated layer outside the model's context window (the finite amount of text a model can process at once). What gets stored, who controls it, and how it stays consistent when multiple agents and users write to the same instance are questions the field hasn't cleanly resolved.

Individual-first versus team-first

Most current enterprise platforms lean individual-first. Microsoft's Copilot, for example, learns a user's role, tone preferences, and working patterns — and stores those as personal memories applied across Microsoft 365 surfaces. That's useful for individual productivity. It doesn't help a team build shared institutional knowledge.

For organizations evaluating agentic platforms, Chintalapani's point is worth taking seriously: the ability to build relational memory retrieval — pulling in relevant context based on what's being asked, not just who's asking — is currently beyond most organizations outside the largest model providers.

That gap is the real story behind the productivity numbers. The agents are learning. They're just not learning for everyone.

Key takeaways

When a worker corrects an AI agent, that correction is typically scoped to their session; colleagues start from zero.
Asana's research found 75% of knowledge workers use AI on the job, but only 5% of companies report measurable productivity gains — a gap the company attributes partly to missing shared context.
In multi-agent workflows, the absence of shared memory can cause agents to contradict each other or repeat work already done by a teammate's agent.
Most enterprise platforms, including Microsoft Copilot, currently take an individual-first memory approach, storing preferences per user rather than per team.
Shared memory architecture is becoming a procurement criterion for engineering and orchestration teams evaluating agentic platforms — not just a technical consideration.

FAQ

What is a shared memory layer in the context of AI agents?

A shared memory layer is a persistent storage system outside the AI model itself that retains corrections, context, and learned preferences — and makes them available to all users on a team, not just the individual who provided the original feedback. Because AI models are stateless by design (they don't retain information between sessions), memory must be built as a separate architectural component.

Why do AI agents forget corrections between users?

Most enterprise AI agents scope their memory to individual sessions or individual user accounts. When one person corrects an agent, that correction is stored only for that person's interactions. A colleague opening the same tool starts with the agent's default behavior, with no access to the previous user's refinements.

What explains the gap between AI adoption and productivity gains?

Asana's research found 75% of knowledge workers use AI on the job, but only 5% of companies report measurable productivity gains. Asana's CPO attributes this partly to the absence of shared enterprise context — agents improve for individuals but don't build institutional knowledge that benefits the whole team. That said, this data comes from Asana, a vendor with a stake in the diagnosis.

How does Microsoft Copilot handle memory?

Microsoft Copilot takes an individual-first approach, learning a user's role, tone preferences, and working patterns and storing those as personal memories applied across Microsoft 365 surfaces. It does not currently offer a team-wide shared memory layer in the same way Asana describes for its platform.

Is shared memory a solved problem?

Not yet. While the concept is well understood, key questions remain open: what context should be stored, who controls it, and how to keep it consistent when multiple agents and users are writing to the same memory instance simultaneously. Experts cited in this reporting say few organizations outside the largest model providers are currently equipped to build robust relational memory retrieval.

Citations

AI agents are learning on the job — just not for your whole team75% of knowledge workers use AI on the job, but only 5% of companies have reported productivity gains, according to Asana's research; Asana CPO Arnab Bose attributed the gap to missing shared enterprise memory context.
VentureBeat — AI and enterprise technology coverageSource publication for reporting on Collate CTO Sriharsha Chintalapani's comments on agent sensitivity to prompt quality and the need for shared memory systems.
AI agents are learning on the job — just not for your whole teamZeta Global CDO Neej Gore described shared context as a 'living memory that compounds intelligence across the enterprise'; Microsoft Copilot described as taking an individual-first memory approach across Microsoft 365 surfaces.