Supermemory has launched Learner-1, a system designed to give any language model the ability to keep improving after deployment by building a running understanding of each user, task, or tenant and injecting relevant tokens into the model's context at inference time. It targets teams building agents that currently reset every session, offering a way to add continual, in-context learning without retraining the underlying model. The launch matters now because agent developers are hitting the ceiling of stateless base models, and Supermemory positions itself as a memory, RAG, user profile, connector, and extractor layer that works with any model .
Under the hood, Supermemory operates as a context API that pairs retrieval with persistent user and task state. It automatically learns from conversations, extracts facts, builds user profiles, handles knowledge updates and contradictions, forgets expired information, and delivers the right context at the right time. On benchmarks the team highlights, Supermemory reaches 95% Recall@15 on LongMemEval while adding only around 720 tokens of context, roughly a 99.4% context reduction , which is the practical claim Learner-1 extends into a general continual learning layer for agents.
The company is led by founder and CEO Dhravya Shah, based in San Francisco, who grew up in Mumbai and built and sold two companies at 16 and 17 before starting Supermemory. The company is developing a universal memory API to solve long-term context limitations in AI applications , and Learner-1 is the next step in that roadmap for founders and infrastructure teams evaluating how to give their agents durable, personalized behavior.
Comments (10)
Sign in to join the discussion.
P
Priya RaghunathanSep 8, 2026
ok wait, so learner-1 actually adapts per agent instead of just doing RAG cosplay? if the eval holds up outside cherry picked demos this is a real unlock.
K
Kostya V.Sep 8, 2026
the landing page has more motion design than my entire product. respect to whoever owns the brand there.
M
Marcus ChenSep 8, 2026
cool announcement but the tweet skipped latency, recall benchmarks, and eviction strategy. memory systems live and die on those three, not on vibes.
L
LenaБроварьSep 8, 2026
who's hiring on the eval side? got a candidate who broke three memory benchmarks last quarter and is finally taking calls.
A
Ada Okafor, PhDSep 8, 2026
curious how learner-1 handles catastrophic forgetting under distribution shift. is there a writeup on the objective or is this closer to a retrieval scheduler with a nicer coat?
H
Hiro TanabeSep 8, 2026
hot take: 90% of agent startups don't need continual learning, they need to stop stuffing 40k tokens of junk into every call. supermemory is quietly solving the latter and calling it the former.
T
the quiet oracleSep 8, 2026
agents without memory are just very expensive autocomplete with commitment issues.
K
Khalid NassarSep 8, 2026
everyone building memory layers is going to get flattened when frontier labs ship native long-term context next cycle. the moat here is a wet napkin.
S
sofia buildsSep 8, 2026
any plans to open source the runtime or at least the eval harness? hard to trust a memory claim I cannot poke at with my own agent.
I
Inesha PatelSep 8, 2026
been watching this team ship since forever, the cadence alone tells you everything. learner-1 feels like the piece that made the rest of the roadmap make sense.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Supermemory
supermemory is building the default engine for memory and continual learning for agents.
supermemory is building the default engine for memory and continual learning for agents.
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Supermemory has launched Learner-1, a system designed to give any language model the ability to keep improving after deployment by building a running understanding of each user, task, or tenant and injecting relevant tokens into the model's context at inference time. It targets teams building agents that currently reset every session, offering a way to add continual, in-context learning without retraining the underlying model. The launch matters now because agent developers are hitting the ceiling of stateless base models, and Supermemory positions itself as a memory, RAG, user profile, connector, and extractor layer that works with any model . Under the hood, Supermemory operates as a context API that pairs retrieval with persistent user and task state. It automatically learns from conversations, extracts facts, builds user profiles, handles knowledge updates and contradictions, forgets expired information, and delivers the right context at the right time. On benchmarks the team highlights, Supermemory reaches 95% Recall@15 on LongMemEval while adding only around 720 tokens of context, roughly a 99.4% context reduction , which is the practical claim Learner-1 extends into a general continual learning layer for agents. The company is led by founder and CEO Dhravya Shah, based in San Francisco, who grew up in Mumbai and built and sold two companies at 16 and 17 before starting Supermemory. The company is developing a universal memory API to solve long-term context limitations in AI applications , and Learner-1 is the next step in that roadmap for founders and infrastructure teams evaluating how to give their agents durable, personalized behavior.
Comments (10)
ok wait, so learner-1 actually adapts per agent instead of just doing RAG cosplay? if the eval holds up outside cherry picked demos this is a real unlock.
the landing page has more motion design than my entire product. respect to whoever owns the brand there.
cool announcement but the tweet skipped latency, recall benchmarks, and eviction strategy. memory systems live and die on those three, not on vibes.
who's hiring on the eval side? got a candidate who broke three memory benchmarks last quarter and is finally taking calls.
curious how learner-1 handles catastrophic forgetting under distribution shift. is there a writeup on the objective or is this closer to a retrieval scheduler with a nicer coat?
hot take: 90% of agent startups don't need continual learning, they need to stop stuffing 40k tokens of junk into every call. supermemory is quietly solving the latter and calling it the former.
agents without memory are just very expensive autocomplete with commitment issues.
everyone building memory layers is going to get flattened when frontier labs ship native long-term context next cycle. the moat here is a wet napkin.
any plans to open source the runtime or at least the eval harness? hard to trust a memory claim I cannot poke at with my own agent.
been watching this team ship since forever, the cadence alone tells you everything. learner-1 feels like the piece that made the rest of the roadmap make sense.