Memory is the moat: State of the Agent market
As models commoditise, lasting advantage for AI agents will come from memory. We explore why knowing what matters, not storing more, decides which agents earn trust, and why the biggest opportunity sits at work.

On this page: 5 sections
Memory is the moat: State of the Agent market
Ask an AI agent to draft a follow-up email and it will do it in seconds. Ask it which follow-up matters this week, to whom, and why, and most agents come up empty.
AI agent adoption is rising fast. Model capabilities are commoditising and execution is likely to follow. As those advantages fade, durable value will come from memory. It is what makes an agent trustworthy enough to act, simple enough to use and specific enough to be worth paying for. It is also what turns an agent into a true System of Action: software that acts on what it knows, rather than on a fixed playbook.
Memory is more than storage
A few years ago, the limit on agents was storage, or how much they could hold about a user or a business. Context windows have since grown dramatically, and the limit has moved. Information storage is now easy; architecting the information that matters is not. Storage is the retention of information an agent can retrieve. Memory now defines the judgement layer on that record. You can think of memory as cognitive layer to decipher which facts still hold and which have been replaced, an active process which learns what matters most from how you work. An agent for work might know my your role, your email tone in emails, your priorities and key stakeholders, but that knowledge is scattered across meeting notes, emails, CRM records and prompts. Not all of it deserves the same weight. Some facts change, some signals conflict and others quietly become irrelevant. The opportunity is not in more storage but in better memory, structured well enough to know what matters now.
Good memory does two things. First, it organises the data a business already holds, so an agent understands what is true today. Second, it learns from how you work:, through the drafts you edit, the nudges you ignore and the decisions you make. This is where defensibility starts. A competitor can copy the model and the integrations, but not months of accumulated feedback.
It's not personal, just business
An agent that spans personal and professional life has obvious appeal, but its potential is limited by what users, and their employers, are willing to share.
The bigger opportunity sits at work, with agents built around how a specific business operates. Work is multiplayer. Deals and projects belong to teams, not individuals. A business-specific agent needs shared memory, clear permissions and an understanding of how colleagues work together, and it gets better as more of the team uses it. Its impact is also easier to measure, whether in deals revived, hours returned or tasks completed. That is what turns a useful assistant into a budget line.
Trust is earned in stages
Access to data can largely be bought. A strong brand, SOC 2 and clear data policies make it more likely that someone will connect their inbox or CRM. Consumers may trade access for convenience, but employers are less willing to take that risk with company information.
Permission to act is different. No certificate will make you comfortable with an agent changing your CRM or emailing a customer on day one. That trust is earned one action at a time, starting with a prep note, then an internal draft, and only later an external message. Auditability helps. An agent that can show what it knows, how it knows it and where it might be wrong is far easier to trust. Doing that well depends on memory that knows when its information has gone stale.
An army of one
The need to brief an agent before every task erodes its value. The better experience is a single assistant for your work, available where you already are, often in mobile messaging apps like iMessage and Whatsapp. Behind the scenes, that assistant can coordinate specialist agents without asking you to manage them. It adds architectural complexity, but it gives you a simpler interface and faster time to value.
Proactivity is also a memory problem. It is easy to mistake activity for usefulness. The best agents may be always on, capturing everything they can, but they only surface what truly matters. For an account executive, that might be a single message: 'A key stakeholder has changed on the stalled deal. I've drafted an email to the new contact. Press to send.' What you act on or dismiss becomes another signal about what you consider important.
Our market map plots the landscape on two axes: personal versus work, and a single agent you simply ask versus builders where you set up an agent for each task. The personal, single-agent quadrant is the most crowded. The top right, where one assistant runs on shared business context, is thinner, and it is where we believe the most durable companies will be built.

The bottom line
Models and integrations will keep getting easier to copy. The context an agent accumulates inside a business, and the trust it earns by correctly executing on that context will not. The winners in this market will not be the agents that do the most, but the ones that know what matters.
In future pieces, we'll explore what an effective memory system needs. If you're building an agent that earns the right to act, or the infrastructure behind it, we'd love to hear from you.
