When someone on the team tweaks the AI agent—better suggestions, better feedback, better context—that improvement disappears the moment a colleague opens the same tool. The fix is not carried over and the next person starts from scratch.
The problem compounds in multi-agent workflows where teams expect agents to share context between users and tasks. Without a shared memory layer, each team member is effectively training a different version of the same agent, and those versions are never synchronized.
This gap is reflected in the numbers. According to Asana’s own research, 75% of knowledge workers use AI in the workplace, but only 5% of companies report increased productivity.
“Model providers are really good at iterating and improving ways of thinking, but what they’re not good at is communicating the business context of the enterprise in a way that people can think of for shared memory,” Asana Chief Product Officer Arnab Bose told VentureBeat.
Asana was built on an agent platform that centralizes context and shared memory. Its Agentic Work Management platform ensures that if any team member edits an agent, that edit applies to everyone on the team.
“That context graph is automatically presented to the agents working in the Asana system, so you don’t need every member of your team to be an expert in operational engineering or context engineering,” Bose said.
said Bose shared memory architecture matters outside of Asana’s own product; this is a design decision that enterprises must make for any multi-agent system.
Shared memory also becomes important as enterprises begin to move from simple single agents to multi-agent workflows that must share context and behaviors.
Memories for multi-agent, multi-platform workflows
Models that power agents are stateless by design, so memory becomes a detached layer outside the context window. While this area of AI innovation is moving toward maturity, the question of what gets stored, who controls it, and how it stays consistent when different agents and users write to the same copy remain largely unsolved.
This can only be managed for use cases with one user. However, in enterprise agent workflows, the idea is for agents to work as a whole team. Most platforms still have agents acting for individuals, leading to duplication of tasks and inconsistent versions of reality and the proliferation of errors. Agents can then also conflict with each other.
Sriharsha Chintalapani, co-founder and CTO of Collate, told VentureBeat in an email that the lack of shared storage is a major bottleneck, especially for multi-agent workflows around sequencing.
"Agents are sensitive to the quality of their instructions," Chintalapani said. "Someone who understands the task well will get more accurate results than someone who is less experienced. This is partly because they can generate more detailed instructions, but also because they can provide better feedback to the agent. The agent remembers the corrections it receives and applies this knowledge to subsequent instructions. The more accurate the feedback, the better the agent will perform for that user. "
He added that organizations should stop treating shared memory as just an operational engineering problem and consider building systems that replicate context in every conversation.
Neej Gore, Zeta Global’s chief information officer, said in a separate email that shared context becomes living memory. "integrates intelligence within the enterprise."
The opportunity may lie in creating AI agents that access memory in a relational way, using relevant context based on what is asked—an approach Chintalapani says that few organizations outside of the largest model providers are equipped to build.
Personal versus team agents
AI agents are already multiplying businesses; simply, many of these act as personal agents that do specific work for individual users. Most prompts start with one person, any file is uploaded by one account, and even for agents that live in a company-wide system, they mostly learn individual user preferences.
Most enterprise AI workflow platforms recognize that memory is important, but approach it from different lenses. For example, Microsoft’s Copilot takes an individual approach by learning the user’s role first within the organization, tone preferences and work patterns are stored as personal memories for later application across the agent’s various Microsoft 365 surfaces.
For engineering and orchestration teams evaluating agent platforms, the question of shared memory is now a procurement criterion – not just technical sophistication. An agent that learns only for the person using it will require ongoing personal maintenance. One who engages in a team-wide memory layer automatically builds institutional knowledge.
