Supermemory launches Learner-1 to make AI agents learn across sessions

Why it matters

Supermemory is betting that the defensible layer in AI agents will be accumulated experience, giving memory providers control over context that persists even when developers swap models.

Dhravya Shah (@DhravyaShah) launched Learner-1 on September 7th, positioning the new Supermemory system as a continual-learning layer that lets AI agents carry experience into later work.

https://x.com/supermemory/status/2097035274094272935

Supermemory introduced Learner-1 in a two-post thread on X, saying it was "doubling down" on memory and in-context learning. The announcement directed developers to the existing Supermemory platform, indicating that Learner-1 is being delivered through Supermemory's context infrastructure rather than presented as a standalone foundation model release.

The release extends Shah's original thesis: capable models still repeat mistakes and relearn user preferences because the surrounding software fails to preserve useful experience. Supermemory sells the storage, understanding and retrieval layer meant to carry that experience between sessions and across agents.

Learner-1 moves memory toward learning

Supermemory's existing system ingests conversations and documents, extracts facts, resolves changing or contradictory information, builds user profiles and retrieves relevant context at request time. Its current product site describes a knowledge graph where facts can be updated, merged, contradicted, inferred or forgotten as new information arrives.

Learner-1 pushes that framing beyond conventional retrieval. A vector database can return previously stored material, while a continual-learning system must identify which parts of experience matter and make them useful during later tasks. Supermemory's announcement ties Learner-1 specifically to in-context learning, placing the work around the model rather than claiming that an agent continuously rewrites its underlying weights.

That distinction matters for developers. External learning can be inspected, edited and moved between models. It also avoids the cost and regression risk of fine-tuning a model whenever an agent encounters a new preference, workflow or failure. Supermemory's existing graph runs at request time, according to its site, allowing developers to change the model behind an agent without discarding the accumulated context.

Supermemory says its API currently processes more than 100 billion tokens per month and returns memories in under 300 milliseconds. Those performance and volume figures are self-reported. The open-source Supermemory repository has accumulated roughly 29,000 GitHub stars, giving Shah a substantial developer distribution channel for the new learning system.

Shah turned a second brain into agent infrastructure

Shah describes himself as a two-time acquired founder who sold products at 16 and 17 before attending Arizona State University. He later worked on Workers AI and other infrastructure projects at Cloudflare, then left college to build Supermemory full time.

Supermemory began as a consumer tool for saving and organizing online material. Demand from developers changed the product. In an October 2025 funding announcement, Shah wrote that businesses were offering to pay for the infrastructure behind the consumer application, prompting Supermemory to expose its memory engine as a service.

Supermemory announced that it had raised $3 million led by Susa Ventures, Browder Capital and SF1.vc, with angels including Cloudflare CTO Dane Knecht, Theo Browne, David Cramer and Julian Weisser. Susa separately described the financing as a $2.6 million pre-seed round, suggesting the larger figure included additional capital or was rounded in Supermemory's announcement.

That money funded a broader context stack spanning memory, retrieval, connectors, document extraction, user profiles and agent filesystems. Learner-1 gives Shah a way to package those components around a larger promise: agents that improve through use instead of merely accumulating searchable transcripts.

Continual learning still fails basic tests

The timing tracks a broader shift in how researchers evaluate agents. The Continual Learning Bench, published in June, measures whether experience from earlier tasks improves an agent's performance on later ones across software engineering, forecasting, database querying and other domains.

Its results left substantial room for improvement. The researchers found that agents often overfit to recent observations or failed to reuse knowledge across tasks. Naive in-context learning even outperformed systems built specifically for memory management in their tests.

Learner-1 is Supermemory's attempt to close that gap from the infrastructure layer. The commercial prize is larger than better chat history. An agent that retains corrections, learns an organization's processes and applies earlier failures to later assignments becomes harder to replace because its useful state compounds inside the memory provider.

That creates the strategic opening Shah is pursuing. Foundation models remain interchangeable for many applications, while accumulated user and business context is specific, persistent and expensive to rebuild. Learner-1 makes Supermemory's bet explicit: the durable part of an agent may sit outside the model.