Investors are paying for agents that learn on the job. The hard part is knowing what they forgot.

On November 2, Trajectory announced a $45 million Series A led by Sequoia at a $300 million valuation, with Nvidia, Bessemer and Bain Capital Ventures participating. The 18-person company connects an open-source model to a customer's application, collects feedback on how it performs, and retrains it. Named customers include Harvey, Rogo, Decagon, Mercor and Nvidia. In August, The Information reported the same Sequoia round at $40 million, following a $15 million seed led by Conviction. The two figures have not been reconciled publicly. Either way, the valuation roughly tripled in a few months. Investors are now paying real money for one promise: agents that get better from their own deployment.

Trajectory is the largest bet in a crowded 2026. NeoCognition, a lab led by Ohio State professor Yu Su, came out of stealth in April with a reported $40 million seed to build agents that specialize through experience. Su told TechCrunch that current agents complete tasks as intended only about half the time. Data trackers attribute the round to an SEC filing, and a lower figure has circulated, but we could not locate a Form D for the company on EDGAR, so treat the size as reported rather than filed. In June, RELAI raised $6.9 million for “verifiable continual learning.” LangChain launched LangChain Labs in May with Harvey, Nvidia, Prime Intellect, Fireworks and Baseten. The memory layer came first: Mem0 raised $24 million in October 2025.

Frontier-lab leaders split on whether this needs new science. Ilya Sutskever told Dwarkesh Patel in November 2025 that he wants to build “a superintelligent 15-year-old” that learns each job after deployment, and he called continual learning “an important and correct thing.” Dario Amodei told the same podcast in February that continual learning “might not be a barrier at all,” and described in-context learning as “kind of like human on-the-job learning, but a little weaker and a little short term.” Google sits between them. Its Nested Learning work and the HOPE architecture treat memory as modules that update at different rates, and the follow-up “Language Models Need Sleep” adds an offline phase that moves new knowledge into parameters. Both are research proposals, not shipping products.

Our thesis is simple. Most of what is sold as “learning on the job” does not change model weights. It changes prompts, memory and routing. LangChain's own framing of continual learning splits it into three layers: model, harness and context. It notes that per-user weight updates are possible in theory but mostly not done in practice, and it calls catastrophic forgetting in weights “an open research problem.” RELAI's fixes, as SiliconANGLE describes them, are prompt changes, tool wrappers, memory updates, workflow adjustments, code repairs and model-routing decisions. Its reported gains, from 39% to 80% on a financial-services agent and from 62% to 96% in a healthcare pilot, are company figures with no published method. Trajectory is the exception that does retrain, but it starts from open-source models, not the frontier models most agents run on. None of this is wrong. It is engineering on the parts of the system that are cheap to change. But it means the word “learning” covers very different mechanisms, with very different failure modes.

The best measurement so far is not flattering. CL-Bench, the Continual Learning Bench from UC Berkeley, Snorkel AI and UW-Madison released in June, tests whether agents improve with experience across six stateful domains, from software engineering to disease-outbreak forecasting. It scores “gain” over each system's own stateless baseline, so prior skill is not counted as learning. The best configuration, full-context in-context learning on Claude Sonnet 4.6, reached only 25.4% normalized gain. The paper's summary is blunt: naive in-context learning outperformed systems built for memory management. Mem0 on GPT-5.4 scored 20.2% gain against 20.1% for plain context on the same model, a tie within confidence intervals. The ACE playbook-memory system was the most expensive configuration, at $62.8 per run, and ranked tenth with 8.6% gain. The authors also report agents overfitting to recent observations or failing to reuse what they learned on new instances.

That last failure is the one to watch. When a deployed agent gets worse after it “learns,” two different things could have happened. It may have lost a skill: a weight update overwrote something, or a memory edit deleted a working rule. Or the skill is still there and the agent picked the wrong one: it retrieved the wrong memory, applied the wrong playbook, or routed to the wrong tool or model. The fixes are opposite. Lost skill calls for protection, replay or rollback. Mis-selection calls for a better router, better retrieval keys, or fewer options. Today's products report one score, so they cannot tell these apart.

Continual-learning theory already separates the two. Kim et al. (NeurIPS 2022) proved that class-incremental learning decomposes into within-task prediction and task-identity prediction, and that strong performance on both is necessary and sufficient. Task-identity prediction is close to out-of-distribution detection. In plain terms, much of what gets called forgetting is a failure to tell which task you are in. Our own Split-MNIST experiments pushed this to the limit. We froze one expert per task, each near-perfect at 0.987, so nothing could be forgotten. Class-incremental accuracy then depended only on picking the right expert. Explicit selection lifted accuracy from 0.169 to 0.635 with no replay buffer. The selector's error rate, though, rose from 0.214 to 0.569 as tasks grew from 2 to 10. The bottleneck moved from memory to choice, and choice got harder as the library grew. That is the situation of an agent with a growing pile of memories, skills and routes.

Research is starting to frame it this way. An October preprint on memory-centric continual learning argues that agent memory should be judged by whether it improves future actions, not by what it retrieves, and proposes separate tests for retention, adaptation and transfer. That is the right direction. What buyers should ask of any “learning” agent is narrower. Which layer changed: weights, prompts, memory or routing? What does it score against plain long context on the same model? And when it fails on a task it used to pass, can it show whether the skill was lost or never selected? Without that last answer, a learning agent can quietly get worse while its memory grows. Our selection-reliability tool computes that verdict from decision logs.