Most AI agents have no memory. Every conversation starts from scratch. That gap matters more than most people realise
Here is something that might sound familiar.
You open an AI assistant at work. You ask it to help with a task — maybe drafting a summary, analysing some data, or pulling together notes for a meeting. It does a reasonable job. You close the window, move on to something else, and come back the next day.
When you return, it has no idea who you are. It does not remember what you worked on together. It does not know what your team decided last week, what project you are in the middle of, or what you asked it to do yesterday. You start from scratch. Every single time.
If that experience sounds frustrating, you have already encountered the problem this post is about — and it may be the most important gap between AI that assists and AI that can be trusted to do real work over time.
Most AI systems in use today are what technologists call stateless. That means they process each interaction independently, with no memory of what came before and no awareness of what might come next. You send a prompt, the system generates a response, and then it forgets everything. The next interaction begins with a blank slate.
Think of it like calling a help desk staffed by strangers. Every time you call, you explain who you are, what the problem is, what you have already tried, and what someone told you last time. The person on the other end might be perfectly capable — but they are starting from zero, and so are you.
Now imagine the alternative. You have a dedicated assistant who knows your history, remembers your preferences, understands what tasks are in progress, and picks up where you left off every time you check in. They do not need you to repeat yourself. They have context. They have continuity. They remember.
That is the difference between stateless and stateful.
A stateful AI system maintains persistent memory, identity, and context across interactions over time.
It sounds simple. But it turns out to be one of the most consequential infrastructure challenges in enterprise technology right now — because stateful infrastructure is the bridge between AI that can help you with a task and AI that can be trusted to do sustained, meaningful work.
I started paying closer attention to this topic after studying the most recent earnings calls from the major cloud and AI infrastructure providers. Several of them — independently and in strikingly similar language — described the shift from stateless to stateful AI as foundational to the next phase of enterprise adoption.
Amazon’s CEO Andy Jassy was the most direct. During the Q1 2026 earnings call, he said the future of using AI models is stateful — that when you are building agents, you cannot start from scratch every time. You need to store identity, store what the agent has done, and store the context it has accumulated. Amazon and OpenAI have committed to co-creating a Stateful Runtime Environment for Amazon Bedrock, specifically designed to address this.
Microsoft is building persistent state through what it calls Work IQ — a layer spanning 17 exabytes of organisational data, including emails, documents, meetings, and conversations, growing 35% year-over-year. CEO Satya Nadella described this as the most important database in any company, constantly changing every second. It is what gives Microsoft’s Copilot the organisational context that makes its responses grounded in real work rather than generic knowledge.
Google introduced “long-running agents” in its Gemini Enterprise Agent Platform, alongside an Agentic Data Cloud designed to give agents persistent access to enterprise context. Their approach makes the data layer the memory layer — so agents can reason over trusted information rather than starting fresh with every query.
And in the enterprise software space, SAP’s recent Knowledge Graph announcement points in a similar direction — agents need a structured understanding of business entities, processes, and relationships if they are going to support mission-critical work where “almost right” is not good enough. That kind of precision requires persistent, structured context that agents can rely on.
The research community is reaching similar conclusions. McKinsey’s AI arm, QuantumBlack, recently published detailed architectural guidance for enterprise agentic platforms. At the centre of their framework is a two-tier memory system: short-term working memory that captures conversation context within a session, and long-term memory that preserves user preferences, past events, and semantic context across sessions. They describe embedding this capability from the start as essential to scaling agent deployments.
Deloitte’s 2026 State of AI in the Enterprise report — based on a survey of more than 3,200 leaders across 24 countries — found that agentic AI adoption is expected to surge from 23% to 74% within two years. But only 21% of companies have a mature governance model for autonomous agents.
That gap — between where adoption is heading and where governance readiness actually is — has a lot to do with the stateless-to-stateful transition. Because governance, as it turns out, depends on memory.
I want to step back from the technical side for a moment, because I think the implications here go well beyond infrastructure.
For anyone who works alongside AI tools today, the stateful shift changes what AI can actually do for you.
A stateless assistant can answer a question. A stateful assistant can manage a project. The difference is not intelligence — both might use the same underlying model. The difference is context and continuity. One forgets. The other remembers. And remembering is what turns a capable tool into a useful partner.
At the organisational level, the implications become even more significant. This is where I keep finding connections to ideas I have been developing in earlier posts.
The more agents act over time, the more memory becomes governance infrastructure.
I have written previously about the governance paradox — the idea that governance is not the obstacle to deploying AI agents at scale, but the condition that makes it possible. And in a recent piece on what I called the AI Payback Question, I explored how governance infrastructure is what makes AI work measurable and provable.
What I had not fully articulated until studying this topic more carefully is that both of those arguments depend on something underneath them. Governance requires persistent state.
You cannot audit an agent that has no memory of what it did. You cannot enforce a policy on an agent that has no persistent identity. You cannot track whether a workflow completed successfully if the agent forgets the workflow existed between sessions. And you cannot prove the economic value of AI work if there is no durable record of the work being done.
Every governance capability that enterprises are asking for — auditability, accountability, compliance, observability — depends on the agent maintaining state. Without it, governance is aspirational. With it, governance becomes implementable.
Deloitte’s research underscores this. The AI risks that companies worry about most all relate to governance — data privacy and security at 73%, legal and regulatory compliance at 50%, governance capabilities and oversight at 46%. Every one of those concerns requires agents that maintain persistent, auditable records of their actions.
Stateful AI is the missing bridge between AI that assists and AI that can be trusted to do work over time. Memory is not just a technical feature. It is the foundation that governance, accountability, and economic value all depend on.
I want to share a few reflections on what I have taken away from this research — not as prescriptions, but as observations that might be useful to others thinking through similar questions.
First, for individuals working with AI tools every day.
If the AI tools you use feel limited — if they cannot seem to understand your context, remember your preferences, or build on previous interactions — the issue may not be the model’s intelligence. It may be the absence of persistent state. As tools evolve toward stateful architectures, the experience of working with AI is likely to change meaningfully. It is worth paying attention to which tools are investing in memory and continuity, and which are still operating on a stateless, session-by-session basis.
Second, for teams evaluating AI platforms.
The question is not just “how capable is this agent?” It is “what does this agent remember — and for how long?” An agent that can reason brilliantly but forgets everything between sessions has a ceiling on the complexity of work it can manage. When evaluating platforms, it may be worth asking how they handle persistent context, long-running workflows, and cross-session memory — because those capabilities determine whether agents can move beyond isolated tasks into sustained, meaningful work.
Third, for leaders thinking about AI governance.
If your organisation is investing in governance frameworks for AI agents — and the research suggests you should be — it is worth understanding that governance depends on stateful infrastructure. Audit trails require persistent records. Policy enforcement requires persistent identity. Compliance requires persistent observability. Before investing heavily in governance policies and frameworks, it may be worth ensuring that your underlying infrastructure can actually support them.
And fourth, for anyone thinking about AI readiness more broadly.
The quality of your organisation’s data — its accuracy, accessibility, and structure — may matter more for AI effectiveness than the choice of model. Stateful agents are only as useful as the context they have access to. An agent with persistent memory but poor underlying data is remembering the wrong things. The organisations that invest in data quality, data governance, and institutional knowledge management are building the foundation that stateful AI needs to be genuinely useful.
The Databricks 2026 State of AI Agents report puts a number to this: companies that use evaluation tools get nearly 6x more AI projects into production, and those that actively practise AI governance get over 12x more. The infrastructure beneath the intelligence is what determines whether agents scale or stall.
The industry has made extraordinary progress on intelligence. What may matter just as much now is continuity — the ability of AI to remember, persist, and build on what it has already done.
Memory — persistent context, identity, and continuity — turns out to be what separates AI that impresses from AI that delivers. And it is the foundation that governance, accountability, and economic value all depend on.
The organisations that recognise this early and invest accordingly — are likely to find themselves in a stronger position than those still focused primarily on model capability alone.
Key Sources
• Deloitte State of AI in the Enterprise 2026 — Report
• Databricks 2026 State of AI Agents Report — Report
• QuantumBlack, AI by McKinsey — “Creating a Future-Proof Enterprise Agentic Platform Architecture” — Article
• QuantumBlack, AI by McKinsey — “Seizing the Agentic AI Advantage” — Report
• SAP Sapphire 2026 — Autonomous Enterprise Announcement
• Deloitte — “AI Agents Are Scaling Faster Than Their Guardrails” — Article
