When someone on a team corrects an AI agent (better prompts, better feedback, better context), that improvement disappears the moment a colleague opens the same tool. The correction is not carried over and the next person starts from scratch.
The problem is exacerbated in multi-agent workflows, where teams expect agents to share context between users and tasks. Without a shared memory layer, each team member effectively trains a different version of the same agent, and those versions are never synchronized.
That gap is noticeable in the figures. According to Asana’s own research, 75% of knowledge workers use AI at work, but only 5% of companies have reported productivity gains.
“Model vendors are getting very, very good at improving reasoning and retry cycles, but what they’re not good at is bringing the enterprise work context into a way where humans can reason for shared memory,” Asana chief product officer Arnab Bose told VentureBeat.
Asana had been building an agentive platform that centers context and shared memory. Its Agentic Work Management platform ensures that if any team member corrects an agent, that correction is applied to all other team members.
“That context graph is automatically provided to agents operating within the Asana system, so there’s no need for every human member of the team to become an expert in rapid engineering or context engineering,” Bose said.
Bose said that shared memory architecture issues beyond the Asana product itself; It is the design decision that companies must make for any multi-agent system.
Shared memory also becomes important as companies begin to move from simple single agents to multi-agent workflows that need to share context and behaviors.
Memories for a multi-agent and multi-platform workflow
The models that drive the agents are stateless by design, so memory becomes a dedicated layer outside of a context window. While this area of AI innovation moves toward maturity, the question of what is stored, who controls it, and how it stays consistent when different agents and users write to the same instance remains largely unresolved.
This is manageable for single-user use cases. However, in enterprise agent workflows, the idea is for agents to work with the entire team. Most platforms have agents that still act for individuals, leading to repetition of tasks and inconsistent versions of reality and the spread of errors. Then the agents could also contradict each other.
Sriharsha Chintalapani, co-founder and CTO of Collate, said in an email to VentureBeat that the lack of shared memory is a major obstacle to multi-agent workflows, especially when it comes to consistency.
"Agents are sensitive to the quality of their indications," Chintalapani said. "Someone with great knowledge of the task will generally obtain more accurate results than someone with less experience. In part, this is because they can create more detailed prompts, but also because they can give the agent better feedback. The agent remembers the corrections it received and applies that knowledge to successive prompts. The more accurate the feedback, the better the agent will perform for that user. "
He added that organizations should stop treating shared memory solely as a quick engineering problem and think about building systems that repeat context in every conversation.
Neej Gore, chief data officer at Zeta Global, said in another email that shared context becomes a living memory that "Composite intelligence across the enterprise."
The opportunity may lie in creating AI agents that retrieve memory relationally, incorporating relevant context based on what is being asked, an approach that Chintalapani says few organizations outside of the largest model vendors are equipped to build.
Personal agents versus team agents
AI agents are already proliferating in companies; The thing is that many of them operate as personal agents who perform specific jobs for individual users. Most prompts start with a single person, all files are uploaded through one account, and even agents that live in a company-wide system mostly learn the preferences of individual users.
Most enterprise AI workflow platforms recognize that memory is important, but approach it through different lenses. For example, the Microsoft copilot. takes an individual-first approach when learning the role of a user within the organization, tone preferences and work patterns, which are then stored as personal memories for the agent to apply to the different surfaces of Microsoft 365.
For engineering and orchestration teams evaluating agent platforms, the issue of shared memory is now a procurement criterion, not just a technical nicety. An agent that learns only for the person using it will require ongoing individual maintenance. One connected to a memory layer of the entire team generates institutional knowledge automatically.
