Convergence raises $12M to build personal AI agents with long-term memory

London startup founded April 2024 by two ML engineers raises a $12M pre-seed led by Balderton for Proxy, an agent that learns admin tasks like a human.

Convergence AI

它在赌什么That useful AI comes from agents that learn like people: Proxy keeps long-term memory and acquires skills across tasks, instead of being built for one workflow at a time.已上线

做的是什么生意

A London-based AI startup building Proxy, a personal AI agent that learns a user's tasks and workflows through long-term memory and continual learning, for workplace and consumer admin tasks.

启动资金：$12M pre-seed led by Balderton Capital, with Salesforce Ventures and Shopify Ventures participating

起因

Convergence was founded in April 2024 by Marvin Purtorab (CEO) and Andy Toulis (CTO), who met at Shopify working on recommender systems and AI assistants before moving to Cohere. Their premise: instead of automating one workflow per product, build a personal agent that learns tasks on the job and remembers them long-term.

经过

In roughly three months the founders assembled a team from Google DeepMind, Meta and PolyAI, then launched Proxy in beta on 24 September 2024 alongside a $12M pre-seed led by Balderton Capital, with Salesforce Ventures and Shopify Ventures participating. Funding was earmarked for the novel models behind Proxy, with a stated focus on continual learning and skill acquisition through memory.

新闻Balderton: $12M pre-seed and Proxy beta launch 新闻Goodwin: advised Convergence on the round

结果

As of the launch, Proxy was in beta with a limited number of spots and a waitlist; Balderton said the product had generated significant interest. The company had not disclosed users, revenue or when the agent would move beyond beta.

背景

Convergence AI was founded in London in April 2024 by Marvin Purtorab and Andy Toulis, machine-learning engineers who met at Shopify working on recommender systems and AI assistants before joining Cohere. Their bet was that the way to genuinely useful AI agents is not another narrow workflow tool but agents that learn like humans — acquiring skills over time through long-term memory.

That premise took concrete form as Proxy, a personal agent designed to learn a user's tasks and workflows, remember them, and extrapolate to new ones. It is powered by what the founders call Large Meta Learning Models, trained to acquire the skill of learning itself, so one agent can cover everything from booking holidays to ordering groceries to workplace admin.

In three months the founders assembled a team from Google DeepMind, Meta and PolyAI and, on 24 September 2024, launched Proxy in beta alongside a $12M pre-seed led by Balderton Capital with Salesforce Ventures and Shopify Ventures participating.

The launch was early: Proxy opened with limited beta spots and a waitlist, no disclosed user numbers or revenue. The bet at that point was that enough people will trust an agent that learns their tasks over time — and that the memory-first model class the founders chose would deliver.

这件事要成立，得有什么

The founders attacked a mechanism — memory — rather than a market, arguing agents plateau because they cannot learn and remember across tasks.
Personal agents were an open field in 2024: Balderton, Salesforce Ventures and Shopify Ventures funded the bet before Proxy had scale.
Team quality was the early signal: ex-DeepMind, Meta and PolyAI hires joined within months of founding, letting two engineers ship beta fast.

可借鉴之处

Bet on a mechanism, not a workflow: Convergence argued agents fail because they lack memory, made that gap the product, and pulled in top VCs and DeepMind-level talent within months.

后续进展

As of 25 September 2024, Proxy was live in beta with a limited number of spots and a waitlist for broader access; Convergence had not disclosed users, revenue or pricing. The $12M pre-seed was earmarked for the models behind the agent, and the company's stated mission was a foundationally different path to useful intelligence.

资料来源

发现哪里写错了？告诉我们。

轮到你了

你刚读完一家。说说你在做什么，看看谁在赌同一件事。

免费账号 · 3 次免费提问 · 不用绑卡