Trajectory raises $40M for AI that learns without forgetting
The morning's biggest bets are on the long game: Sequoia is funding a team trying to make models keep learning after deployment, Beijing's order to unwind Meta's Manus deal is nearly done, and agent observability keeps pulling in money.
Trajectory, a startup founded by former DeepMind, Apple, OpenAI, and Meta staffers, has raised $40 million led by Sequoia at a $300 million valuation to build continual learning models, according to The Information. Continual learning is the field's answer to one of its most stubborn problems: today's models freeze at deployment, and fine-tuning them on new knowledge tends to wipe out old skills — a failure mode known as catastrophic forgetting. Chinese tech outlet 36Kr has profiled Trajectory building its platform on Thinking Machines' Tinker fine-tuning stack with a self-distillation technique that lets a model act as its own teacher, picking up new skills from demonstrations without eroding what it already knows. That positions the startup less as another pre-training lab and more as an infrastructure play for "continual learning as a service" — frequent, on-demand model updates without the forgetting tax. Sequoia's willingness to pay a $300 million valuation on a problem Andrej Karpathy has estimated could take a decade to crack says the market believes learning-while-deployed is the next capability frontier after reasoning and agents; the risk is that the field is crowded with divergent approaches, from external memory to weight edits, and nobody has proven which one wins at scale.
Manus says it will "soon return to operating as an independent company," signaling it is close to finalizing the reversal of its acquisition by Meta, The Information reports. The unwind traces back to April, when Beijing's NDRC ordered Meta to undo its $2 billion-plus purchase of the Singapore-incorporated agent startup — the first use of China's foreign-investment security review to reverse a completed cross-border AI deal — and by June Meta had reportedly cut Manus off from its internal systems. A finalized separation would make Manus one of the few companies to survive a Big Tech acquisition being forcibly unwound, and it lands the general-purpose agent startup back in the race as an independent player. Watch for what Manus keeps — and what stays with Meta — once the split is officially complete.
FriskAI launched with $3.6 million in pre-seed funding, led by MaC Venture Capital, to give enterprises a record of what their AI agents actually do in production. The Los Angeles startup's runtime-intelligence platform sits alongside agents, logs every tool call with its arguments and timing, builds a behavioral profile per task and tool, and flags when a new deployment starts acting differently from the last one — no preset rules required. With SDKs for Python and TypeScript and adapters for LangChain, the Claude Agent SDK, and Strands, it's aimed squarely at healthcare, insurance, and financial services teams that must explain an agent's decisions to auditors; Sana Benefits is an early user. The pitch lands in a category that is suddenly well-funded — Zenity announced a $125 million round this month for agent security — and FriskAI's bet is that "what did the agent actually do?" becomes a compliance question before it becomes an engineering one.
What to watch: whether Trajectory publishes benchmarks that prove its self-distillation approach holds up against catastrophic forgetting at scale.
Continual learning has been called the hardest unsolved problem in AI — do you think a startup can crack it before the big labs do? Tell us in the comments.
Sources: Techmeme · 36Kr · The Information via Techmeme · Resultsense · SiliconANGLE · FriskAI
