When someone on a team improves an AI agent – better signals, better feedback, better context – that improvement disappears the moment a colleague opens the same tool. The improvement does not transfer, 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 never sync.
That difference shows up in the numbers. According to Asana’s own research, 75% of knowledge workers use AI at work, but only 5% of companies have reported productivity gains.
“Model providers are getting really good at improving logic and retry, but they’re not good at bringing enterprise task context to the table in a way that humans can reason about shared memory,” Arnab Bose, chief product officer at Asana, told VentureBeat.
Asana was building toward an agentic platform that puts context and shared memory at the center. Its agentive task management platform ensures that if a team member makes a correction to an agent, that correction applies to everyone else on the team.
“That context graph is automatically provided to agents working inside Asana’s system, so you don’t have to have every human member of the team become an expert in prompt engineering or context engineering,” Bose said.
Bose said the shared memory architecture has implications beyond Asana’s own product; This is the design decision that enterprises need to 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 need to share context and behavior.
Memories for multi-agent, multi-platform workflows
Model powering agents are stateless by design, so memory becomes a dedicated layer outside a context window. While this area of AI innovation is moving toward maturity, the questions of what is stored, who controls it, and how it remains consistent when different agents and users write to the same instance remain largely unresolved.
This is manageable for just one user use cases. However, in enterprise agentic workflows, the idea is for agents to work with the entire team. Most platforms have agents who still work for individuals, leading to repetitive tasks and the spread of inconsistent versions of reality and inaccuracies. Agents could then also contradict each other.
Sriharsha Chintlapani, co-founder and CTO of Collate, said in an email to VentureBeat that the lack of shared memory is a major hurdle for multi-agent workflows, especially for stability.
"Agents are sensitive to the quality of their signals," Chintlapani said. "Someone with a deep understanding of the task will generally achieve more accurate results than someone less experienced. Partly this is because they are able to make more detailed signals, but also because they are better able to respond to the agent. The agent remembers the improvements it receives and applies that knowledge to successive signals. The more accurate the feedback, the better the agent will perform for that user. "
He said organizations should stop treating shared memory as just a quick engineering problem and think about building systems that replicate context in every interaction.
Neese Gore, chief data officer at Zeta Global, said in a separate email that the shared context becomes a living memory. "Blends intelligence across the enterprise."
The opportunity may lie in building AI agents that retrieve memory relatively, pulling in relevant context based on what is being asked – Chintalapani says few organizations outside the largest model providers are equipped to create this.
Individual vs Team Agent
AI agents are already proliferating enterprises; It’s just that many of these act as personal agents performing specific tasks for individual users. Most signals start with one person, any file uploaded by an account, and even for agents living in a company-wide system mostly learns individual user preferences.
Most enterprise AI workflow platforms agree that memory is important but look at it through different lenses. For example, Microsoft’s CoPilot takes an individual-first approach by learning the user’s role, tone preferences, and working patterns within the organization, which are then stored as personalized memories for the agent to apply across various Microsoft 365 surfaces.
For engineering and orchestration teams evaluating agentive platforms, the shared memory question is now a purchasing criterion – not just a technical nuance. An agent that learns only for the person using it will require constant personal maintenance. The individual connected to the team-wide memory layer automatically creates institutional knowledge.
<a href