When an individual on a team refines an AI agent’s performance through better prompts, feedback, or context, that learning is ephemeral. It vanishes the moment a colleague accesses the same tool, forcing them to start from square one. This lack of persistent, shared learning becomes particularly problematic in multi-agent workflows, where teams expect AI agents to maintain context across users and tasks. Without a unified memory layer, each team member effectively trains a distinct version of the same agent, leading to a perpetual state of divergence.
This inefficiency is reflected in operational metrics. Asana’s internal research indicates that while 75% of knowledge workers utilize AI in their roles, only a scant 5% of companies report tangible productivity gains. Arnab Bose, Asana’s Chief Product Officer, highlighted this disconnect in a conversation with VentureBeat: “Model providers are becoming exceptionally adept at enhancing reasoning and retry loops, but they are struggling to integrate enterprise work context in a manner that facilitates shared, human-understandable memory.”
Asana has been strategically developing an agentic platform designed around the principles of context and shared memory. Its Agentic Work Management platform ensures that any improvement made by one team member to an AI agent is universally applied across the team. Bose elaborated, “This contextual graph is automatically supplied to agents operating within Asana’s ecosystem, obviating the need for every human team member to become an expert in prompt or context engineering.” He further stressed that this shared memory architecture is a critical design decision for enterprises embarking on any multi-agent system, extending beyond Asana’s specific product.
The imperative for shared memory becomes even more pronounced as organizations transition from basic, single-agent applications to complex multi-agent workflows that demand seamless context and behavior sharing.
Memory Systems for Cross-Platform, Multi-Agent Workflows
The underlying models for AI agents are inherently stateless, necessitating a distinct memory layer that operates independently of the limited context window. While this domain of AI innovation is rapidly maturing, fundamental questions persist regarding what data is stored, who governs access, and how consistency is maintained when multiple agents and users interact with the same memory instance. These challenges are generally manageable for single-user scenarios.
However, enterprise agentic workflows are predicated on collective team functionality. Currently, many platforms feature agents that operate primarily on an individual level, leading to task redundancy, inconsistent outputs, and the propagation of errors. This can also result in agents contradicting one another.
Sriharsha Chintalapani, co-founder and CTO of Collate, identified the absence of shared memory as a significant impediment to the efficacy of multi-agent workflows, particularly concerning consistency. He explained to VentureBeat via email, “Agents are highly sensitive to prompt quality. Individuals with a deep understanding of a task typically achieve more accurate results than less experienced users. This is partly due to their ability to construct more detailed prompts and provide superior feedback. The agent internalizes corrections, applying that learned knowledge to subsequent prompts. The more precise the feedback, the better the agent performs for that specific user.”
Chintalapani advocates for organizations to reframe shared memory not merely as a prompt engineering challenge, but as a foundational architectural consideration for systems designed to replicate context across all interactions.
Neej Gore, chief data officer at Zeta Global, echoed this sentiment, describing shared context as a dynamic repository that “compounds intelligence across the enterprise.”
The future opportunity may lie in developing AI agents capable of relational memory retrieval, dynamically accessing relevant context based on the query. Chintalapani notes that few organizations, outside of major model providers, possess the capabilities to implement such an approach.
Distinguishing Personal Agents from Team Agents
AI agents are already widely deployed within enterprises, though many function as personal assistants, executing tasks tailored to individual users. Typically, prompts originate from a single person, file uploads are linked to one account, and even agents integrated into company-wide systems often learn individual user preferences.
Most enterprise AI workflow platforms acknowledge the importance of memory but approach its implementation from different perspectives. Microsoft’s Copilot, for instance, adopts an individual-centric model. It learns a user’s role, stylistic preferences, and work habits, storing these as personal memories that the agent applies across various Microsoft 365 applications.
For engineering and orchestration teams evaluating agentic platforms, the availability of shared memory is evolving from a desirable feature to a key procurement criterion. An agent limited to learning for a single user necessitates continuous, individualized maintenance. Conversely, an agent integrated with a team-wide memory layer automatically cultivates institutional knowledge.
Business Style Takeaway: The lack of persistent, shared memory in current AI agent deployments is a critical bottleneck hindering enterprise-wide productivity gains, despite widespread adoption. Solutions that establish a unified, team-accessible memory layer are essential for unlocking the true collaborative potential of AI and building scalable, institutional intelligence.
Based on materials from : venturebeat.com