AI agents learn on the job, but not for your entire team

When someone in the team corrects artificial intelligence agents — better tips, better feedback, better context — this improvement disappears when colleagues open the same tool. The amendment does not transfer, the next person starts from scratch.

The issue has become more complex in the multi-agent workflow, and the team hopes that the agent will be able to share context across users and assignments. Without shared memory, each team member effectively trains different versions of the same agent, which will never be synchronized.

This gap is reflected in figures. According to Asana's own research, 75 per cent of intellectuals use artificial intelligence at work, but only 5 per cent of companies report increased productivity.

Arnab Bose, Chief Product Officer of Asana, told Venture Beat: “Model providers are indeed very good at improving reasoning and re-trying cycles, but they are not very good at providing the business environment in a way that humans can reason to share their memory.”

Asana has been building a proxy platform centred on context and shared memory. Its proxy work management platform ensures that if any team member corrects the agent, the correction applies to all other team members.

“This context diagram will automatically be provided to the agent operating in the Asana system, so you do not have to let every member of the team become an expert in instant or context works,” said Bose.

Bose states that shared memory structures are important beyond Asana ' s own product range. This is the design decision that an enterprise needs to make for any multi-agent system.

Sharing memory is also important when an enterprise begins to move from a simple single agent to a multi-agent workflow that requires sharing context and behaviour.

Multi-agent, multi-platform workflow memory

Models that support agents are designed to be non-state, so memory becomes a dedicated layer outside the context window. While the area of artificial intelligence innovation is maturing, the question of what to store, who to control it and how to maintain consistency when different agents and users write the same example remains largely unresolved.

This is manageable for one user. However, in the business proxy workflow, the idea is to work with the entire team. Most of the platforms are still represented by individuals, which can lead to duplication of tasks, inconsistent versions of reality and erroneous dissemination. Agents can also contradict each other.

Collat e Co-founder and Chief Technical Officer Sriharsha Chintalapani, in an e-mail to Venture Beat, stated that lack of shared memory was a major obstacle to multi-agent workflows, especially in terms of consistency.

“Agents are sensitive to the quality of the alert,” Chintarapani said. “People with in-depth knowledge of the mandate usually obtain more accurate results than those with less experience. This is partly due to their ability to construct more detailed tips, but also because they can provide better feedback to agents. The agent will keep in mind the corrections received and apply them to subsequent reminders. The more accurate the feedback, the better the agent will perform for the user.”

He added that organizations should not simply view shared memory as a matter of immediate engineering, but should consider constructing a system that repeated context during each dialogue.

Zeta Global Chief Data Officer Neej Go.In another e-mail, it is stated that sharing context becomes “living memory” and “enhances the intelligence of the enterprise as a whole”.

The opportunity may be to build an artificial intelligence agent capable of retrieving the relevant memory, extracting the context based on what was asked - Chintalapani states that, apart from the largest model provider, few organizations have the capacity to construct this approach.

Personal and Team Agent

Artificial intelligence agents have proliferated in enterprises; only many of them operate as personal agents and work specifically for individual users. Most tips begin with one person, and any document is uploaded from an account, and even for agents living in company-wide systems, most of them know the preferences of individual users.

Most enterprise AI workflow platforms recognize the importance of memory but view it from different perspectives. For example:Microsoft Copilot uses a personal priority approach, by understanding the role, tone and mode of work of the user in the organization, and then storing it as a personal memory for the agent to apply on a different Microsoft 365 interface.

The issue of shared memory is now a procurement standard, not just a technical detail, for the engineering and programming team to assess the proxy platform. Only agents who learn to use it will require sustained personal maintenance. A system connected to a team-wide memory automatically builds institutional knowledge.