Reload builds shared memory for AI agents

One of the biggest flaws of today's AI agents is their amnesia. Every new session starts from a blank slate, and every agent lives in its own information bubble, unaware of what its "colleagues" are doing. Startup Reload has set out to attack exactly this problem: the company raised $2.275 million in a round led by venture fund Anthemis and simultaneously launched its first AI employee, an agent named Epic. The product is built on the idea of shared memory, which lets different agents keep context, exchange knowledge and work in sync, the way people do on a healthy team.

To see why this matters, it helps to look at how current agentic systems work. Companies such as OpenAI, Anthropic and Google have taught their models to carry out complex multi-step tasks: browsing web pages, writing code, managing files, calling APIs. But when several agents have to work as a team, the architecture comes apart at the seams. An agent that gathered data on competitors in the morning no longer remembers by evening what exactly it found. The agent responsible for writing the report has no idea what its "colleague" concluded. Every tool stands alone, and every session is a separate island. Reload saw an opportunity in exactly this gap.

The company's architectural answer is a centralized memory layer that agents query as a shared knowledge store. Technically, it resembles RAM for a team: one agent writes a fact there, another reads it and uses it in its own task, a third adds to it, all without having to pass huge volumes of context through prompts each time. This approach tackles several problems at once: it lowers compute costs, speeds up the agents and removes contradictions that arise when different parts of a system operate on different versions of the same information.

The first product on this architecture, the Epic agent, is aimed at business processes, although the company is so far revealing specific use cases only in small doses.

The choice of Anthemis as lead investor says a lot. The fund specializes in fintech and insurance, sectors where data continuity, decision auditing and process coordination are critical. It is not a random choice: financial organizations run hundreds of operational agents that badly need a shared "working memory." If Reload manages to establish itself in this niche, the potential for scaling is enormous, since the financial sector traditionally pays well for reliable infrastructure. The $2.275 million is small by AI-industry standards, where rounds often run to hundreds of millions, but for an early-stage infrastructure startup it is enough to validate the hypothesis and reach the first enterprise customers.

Reload's arrival fits into a broader debate about the architecture of multi-agent systems. Nearly all the major players are part of it today: Microsoft is integrating agent capabilities into Copilot, Google is building an ecosystem around its Gemini models, and hundreds of startups are trying to claim specialized niches in this stack. The key question is who will own the "nervous system" of the agent-driven enterprise. Reload is betting that memory is the layer that will prove the most valuable and the hardest for competitors to replicate. The logic is convincing: switching a model is easy, but switching a memory store that has accumulated months of working context is a task on a fundamentally different scale.

For end users and companies adopting AI, the success of products like this means a qualitative shift in how automation is perceived. An agent that remembers your preferences, understands a project's history and knows what its "partner" did an hour ago is no longer just a tool but something approaching a full-fledged member of the workforce. The line between software and an "employee" is blurring, and infrastructure companies like Reload help determine how real that metaphor turns out to be.

Reload remains a small player with an ambitious idea and, by industry standards, modest funding. But if the company can prove that shared memory really makes agentic systems more reliable and efficient, it risks becoming one of those unglamorous but indispensable components that every serious AI infrastructure ends up needing. In an era when everyone is racing for the smartest model, the winner is the one who builds what makes those models remember.