Why your AI agents keep learning alone instead of as a team

AI agents learn individually rather than collectively

AI adoption is moving fast. According to Asana’s research, 75% of knowledge workers already use AI at work. But only 5% of companies report measurable productivity gains. That gap deserves attention because it suggests the technology is not the main limitation. The way companies deploy it is.

Today, most AI agents learn at the individual level. One employee discovers a better prompt, corrects a mistake, or provides richer context. The agent becomes more useful for that person. Then another employee opens the same tool and starts from the beginning. The previous improvement is gone because it was never shared.

This creates a hidden cost that many organizations underestimate. Teams repeat the same work. Different employees receive different answers for similar tasks. Best practices stay with individuals instead of becoming organizational knowledge. As AI becomes part of daily operations, this fragmented learning limits the return on investment.

For executives, the question is no longer whether employees use AI. They already do. The question is whether every interaction makes the entire organization smarter or only helps one employee. Companies that solve this problem will improve faster because knowledge compounds across teams instead of remaining isolated.

Arnab Bose, Chief Product Officer at Asana, summarized the challenge in an interview with VentureBeat. He said that model providers have become “really, really good at improving reasoning and retry loops,” but they are “not good at bringing the enterprise work context in a way that human beings can reason about for shared memory.”

That distinction matters. Better reasoning makes an AI agent more capable. Shared organizational memory makes it more valuable inside a business. These are different problems, and both need to be addressed.

Shared memory is critical for effective multi-agent workflows

The next stage of enterprise AI is not one intelligent agent. It is many agents working together across departments, applications, and business processes. That only works if they operate from the same source of knowledge.

Without shared memory, every agent develops its own understanding of the business. One agent may have updated information while another uses outdated context. One team may improve an AI workflow, while another unknowingly repeats the same corrections. The result is inconsistent decisions, duplicated effort, and avoidable mistakes.

A shared memory layer changes this. Instead of storing knowledge with individual users or isolated agents, it creates a common repository of approved context, corrections, and business knowledge. Every agent can access the same information, allowing improvements made by one team to benefit the rest of the organization automatically.

This is especially important as companies automate more complex workflows. Finance, legal, sales, customer support, and engineering increasingly depend on connected AI systems rather than standalone assistants. If these systems cannot share reliable context, they cannot consistently support business operations at scale.

Arnab Bose, Chief Product Officer at Asana, argues that this is not simply an Asana product feature. He sees shared memory as a fundamental architectural decision that every enterprise will eventually have to make when deploying multi-agent AI systems.

Sriharsha Chintalapani, Co-founder and CTO of Collate, reinforced this point. He told VentureBeat that the lack of shared memory is a major obstacle for multi-agent workflows because consistency depends on more than model quality. He explained that agents are highly sensitive to prompt quality. Employees with deeper expertise usually produce better prompts and better feedback, leading to stronger results for themselves. However, unless those improvements become shared organizational knowledge, the company continues to depend on individual expertise instead of building repeatable capability.

For business leaders, this changes how AI platforms should be evaluated. Model performance will always matter, but it is no longer enough. Procurement decisions increasingly need to consider how knowledge is captured, governed, and shared across teams. An AI platform that learns once and benefits everyone creates a very different business outcome from one that requires every employee to continuously train their own assistant.

Asana’s agentic platform uses shared memory to turn individual improvements into organizational knowledge

Many companies recognize that AI needs context. The difference is how that context is managed. Asana’s approach is to make context a shared organizational asset rather than something that remains with individual users.

Its Agentic Work Management platform is built around a shared context graph. When one team member improves an agent by correcting an answer, providing better instructions, or adding important business context, that improvement becomes available to other users working within the same environment. The objective is to ensure that the system improves continuously as more people use it.

This changes how organizations scale AI. Instead of asking every employee to become skilled at prompt engineering, the platform provides relevant organizational context automatically. Employees can focus on the work itself while the system supplies the information the agent needs to perform consistently.

Arnab Bose, Chief Product Officer at Asana, explained this approach in an interview with VentureBeat. He said, “That context graph is automatically provided to agents operating inside Asana’s system so you don’t have to have every human member of the team become an expert at prompt engineering or context engineering.”

For executives, this is an important operational consideration. AI adoption often slows when organizations expect employees to learn new technical skills before they can benefit from the technology. A platform that distributes context automatically reduces that barrier and allows organizations to expand AI usage across departments without creating additional complexity for users.

There is also a governance advantage. When context is managed centrally, organizations have greater visibility into the knowledge being shared with AI systems. This creates opportunities to standardize processes, improve consistency, and reduce variation between teams. Over time, the organization develops institutional knowledge that remains available even as employees change roles or leave the company.

The broader lesson extends beyond Asana. Shared memory should be viewed as part of enterprise infrastructure rather than an application feature. As organizations deploy more AI agents across business functions, the ability to preserve and distribute organizational knowledge will become increasingly important.

Building reliable shared memory remains one of enterprise AI’s biggest technical challenges

The need for shared memory is becoming clearer, but implementing it is far from simple. Today’s large language models are stateless by design. They process the information they receive during a conversation, but they do not permanently retain knowledge unless an external memory system stores it.

This means enterprises must build a dedicated memory layer outside the model. That layer determines what information is stored, how it is updated, who can modify it, and how conflicts are resolved when multiple users or AI agents contribute new information. These decisions directly affect the quality, reliability, and trustworthiness of AI-generated outputs.

The challenge becomes much larger in enterprise environments. A single user interacting with one AI agent creates relatively few coordination issues. A business operating hundreds or thousands of agents across multiple departments faces a very different situation. Different agents may access different data sources, receive conflicting updates, or produce inconsistent recommendations if they are not connected through a reliable memory architecture.

Sriharsha Chintalapani, Co-founder and CTO of Collate, told VentureBeat that the absence of shared memory remains a major obstacle to achieving consistency in multi-agent systems. His observation reflects a broader industry challenge. The technology for reasoning has advanced quickly, but the infrastructure for maintaining accurate, shared organizational knowledge is still evolving.

Another important consideration is governance. Enterprise memory should not become an uncontrolled collection of information. Organizations need clear policies for data quality, permissions, security, and retention. Sensitive business information must remain protected while still being available to authorized AI systems that require it to perform useful work.

For business leaders, these architectural decisions should receive the same level of attention as model selection. The strongest AI model cannot consistently deliver reliable business outcomes if it operates on incomplete, outdated, or inconsistent organizational knowledge. As multi-agent systems become more common, the quality of the shared memory layer will increasingly determine whether AI delivers isolated productivity gains or sustained enterprise-wide value.

Organizations need to move beyond prompt engineering and build systems that preserve shared context

Many organizations still approach AI performance as a prompt engineering challenge. They invest time teaching employees how to write better prompts or provide more detailed instructions. While this can improve results, it does not solve the larger problem. Every improvement remains tied to the individual unless the organization has a way to capture and reuse that knowledge.

This creates uneven performance across the business. Employees with deeper domain expertise usually produce better outputs because they know how to ask better questions and provide more useful feedback. New employees or those with less experience often receive weaker results, even when using the same AI platform.

Sriharsha Chintalapani, Co-founder and CTO of Collate, explained this issue in comments to VentureBeat. He said, “Agents are sensitive to the quality of their prompts. Someone with a strong understanding of the task will generally get more accurate results than someone less experienced.” He added that better results come from stronger prompts and from higher-quality feedback, allowing the agent to improve over time for that specific user.

Chintalapani argued that organizations should stop treating shared memory solely as a prompt engineering problem and instead build systems that preserve context across every conversation. This changes the objective from helping individuals become better AI users to helping the organization continuously improve its collective knowledge.

One promising direction is relational memory retrieval. Instead of retrieving all available information, AI agents identify and access the most relevant organizational knowledge for the task at hand. This reduces unnecessary information while improving the quality and consistency of responses. According to Chintalapani, few organizations outside the largest AI model providers currently have the capability to build these systems effectively.

Neej Gore, Chief Data Officer at Zeta Global, described the broader opportunity in comments to VentureBeat. He said that shared context becomes a living memory that “compounds intelligence across the enterprise.”

For executives, this is an important strategic shift. Competitive advantage will increasingly come from how well an organization captures, governs, and applies its own knowledge. Prompt engineering remains useful, but it should complement a broader knowledge strategy rather than replace it. Companies that invest in shared context can reduce duplicated work, improve consistency, and accelerate learning across teams.

Enterprise AI is shifting from personal assistants to organization-wide intelligence

Most AI agents used in businesses today are still designed around individuals. They learn a user’s writing style, work preferences, documents, and daily habits. This makes them more useful for personal productivity, but it does little to improve how teams work together.

This individual-first approach reflects the current design of many enterprise AI platforms. A user uploads files, interacts with the agent, and gradually improves its performance through repeated use. Those improvements generally remain attached to that user rather than becoming part of the organization’s shared knowledge.

Microsoft Copilot is an example of this model. Copilot learns information such as a user’s role within the organization, preferred tone, and working patterns, storing these as personal memories that are used across Microsoft 365 applications. This creates a more personalized experience, but the memory primarily benefits the individual rather than the wider team.

As organizations expand AI adoption, this design is becoming a business decision rather than simply a technical one. AI agents are increasingly expected to support cross-functional work involving finance, operations, sales, engineering, customer service, and other business units. These workflows require shared organizational understanding.

This is why shared memory is becoming an important procurement criterion for enterprise AI platforms. Decision-makers need to evaluate more than model accuracy or feature lists. They also need to understand how a platform captures organizational knowledge, how that knowledge is governed, and whether improvements made by one team can benefit the entire enterprise.

An AI agent that only learns from one employee requires continuous individual maintenance. A platform that builds and distributes institutional knowledge creates value that grows with every interaction. Over time, that allows organizations to improve consistency, reduce repeated work, and make AI a core part of how teams operate rather than simply another personal productivity tool.

For business leaders, the direction is becoming clearer. The next phase of enterprise AI is not defined by having more AI agents. It is defined by creating systems where those agents learn together, operate from the same trusted knowledge, and improve continuously as the organization evolves.

Key highlights

• Build AI that learns for the whole organization: Individual AI improvements do not automatically benefit the rest of the business, limiting productivity gains. Leaders should prioritize platforms that capture and reuse knowledge across teams instead of isolating learning to individual users.
• Make shared memory a core architecture decision: Multi-agent AI cannot scale reliably if every agent operates with different context. Evaluate AI platforms on how they maintain consistent, shared organizational knowledge.
• Turn AI interactions into institutional knowledge: Systems that automatically share corrections and business context reduce duplicated effort and improve consistency. Leaders should look for solutions that lower the need for prompt engineering while strengthening enterprise-wide knowledge over time.
• Treat memory governance as a strategic capability: Shared memory requires clear rules for data quality, access, and consistency because AI models are inherently stateless. Strong governance is essential to ensure reliable outputs as AI deployments expand across the business.
• Shift investment from prompts to knowledge systems: Better prompts improve individual results, but shared context improves organizational performance. Leaders should invest in systems that preserve, retrieve, and apply enterprise knowledge across conversations and workflows.
• Buy for enterprise intelligence: Many AI platforms optimize for individual users, but enterprise value comes from shared organizational learning. Make shared memory and institutional knowledge key procurement criteria when selecting AI platforms.