Engram raises $98M to teach AI models your company, not the open web

Engram, a 13-person AI lab building models that learn from a single company's private context instead of the open web, has raised $98 million at a reported $600 million valuation. The round was led by General Catalyst, with Kleiner Perkins, Sequoia, Factory, Modern, Amplify, Neo, and SV Angel joining, and advisors including Wiz CEO Assaf Rappaport, former OpenAI researcher Andrej Karpathy, and UC Berkeley professor Pieter Abbeel.

The premise is a direct challenge to how the largest AI labs have been scaling. Instead of spending more compute to train on more public text, Engram retrains a strong pre-trained model on a customer's own knowledge base. The lab claims that, on internal data, its models can be 10x to 100x more token-efficient than a frontier general-purpose model, because the model no longer has to rediscover the company every time it answers a question. Engram is already running this loop on its own company data every day, with a target of every hour, and eventually every minute. The same gap that pushed AWS to ship its own Continuum and Context layer for agents is the gap Engram is raising against, from the other side of the stack.

For more on the underlying infrastructure, see our AI Infrastructure resource page.

Why the company is betting on memory, not more parameters

Engram's founders are veterans of the context-compression, retrieval, LoRA, and long-context research communities. The team's argument, laid out in a public launch post on engram.com, is that the next scaling axis for AI is not parameter count, it is per-customer compute. Public pretraining has plateaued in usefulness for the kinds of questions an enterprise actually asks, because most of the answer lives inside a private corpus the public model has never seen. The standard fix today is to stuff the model with retrieval, which works but costs tokens on every call and forgets what it learned the moment the chat window closes.

Engram's pitch is to spend training compute, not inference compute, on the customer's context. The lab runs a continual training job that updates a base model against the customer's Slack, GitHub, Notion, documents, and other internal systems, so the resulting model has already absorbed the company before the user even types a prompt. That is a different cost curve from retrieval. It also means the model is sticky in a way an enterprise deployment of a frontier API is not. If your model is the only one that knows your company, switching costs rise sharply.

The team frames this as a path to models that feel like a teammate rather than a search box. In their words, the north star is a single training algorithm that can absorb arbitrary amounts of data into a model that gets continually better, with the retraining cadence shortening over time. Today's architecture runs the full retraining loop daily, with hourly and per-minute targets on the roadmap.

Design partners, and what each one is testing

Engram is not selling a chat product. The first surface is an API for agents that learn on very large shared knowledge workspaces, with three named design partners. With Notion, the team is building custom agents that understand entire Notion workspaces. With Harvey, the legal AI company, Engram is developing models that internalize the knowledge of an entire law firm and can search across many client matters. With Microsoft, the team is piloting Engram models inside M365 to deliver cost-efficient, customized agents for enterprise customers.

The partner list is short and deliberate. Notion brings one of the richest user-generated knowledge graphs on the public internet and a fast-moving product team. Harvey brings a hard evaluation surface, because lawyers will catch a model that has not actually read a brief. Microsoft brings the deployment path that matters most for enterprise reach, since M365 is where most of the corporate knowledge already lives. The fact that Engram is shipping through an API rather than a competing chat app also signals that the lab expects the value to accrue to whoever owns the agent, not the model.

The investor signal, and what to watch next

The round reads as a bet that enterprise AI will split into a long tail of vertical and customer-specific model providers, rather than consolidating around a few frontier APIs. General Catalyst, Kleiner Perkins, and Sequoia are not known for funding tiny research teams with no near-term revenue plan. The fact that all three are on the cap table, alongside neo-funds like Factory and Modern, suggests investors see Engram as a platform play, not a product bet. The 13-employee headcount and $600 million valuation, reported by CTech and other outlets covering the round, also stands out. The implied per-employee valuation is unusually high, which usually means investors are pricing the team and the research direction, not the revenue. That is consistent with Engram's framing of itself as a research lab that ships a product, not the other way around. The advisors reinforce the read. Assaf Rappaport built Wiz into one of the fastest-growing enterprise security companies of the last decade. Andrej Karpathy is a former OpenAI co-founder and Tesla AI director. Pieter Abbeel runs one of the most cited robotics and reinforcement learning groups in academia. The cap table and the brain trust are pointing in the same direction.

Three signals will tell us whether Engram's memory-first bet is real. First, whether the Notion integration ships as a customer-facing product rather than a closed design partnership, because that is the first moment an outside user can actually feel the difference. Second, whether the Harvey relationship moves from a handful of firms to a wider rollout, because legal knowledge is the cleanest stress test for a model that claims to know a specific company. Third, whether Microsoft exposes Engram models through a labeled SKU inside M365, because the path to recurring enterprise revenue runs through procurement, and procurement needs a line item.

If those three signals land, the broader question is whether the rest of the field has to follow. Frontier labs have already started adding long-term memory features to consumer products, but those are retrieval layers on top of a static base model, which is the architecture Engram is arguing against. A research lab that can show, in production, that a 100x-token-efficiency model is good enough on enterprise tasks, would force a real shift in how compute budgets are split between pretraining and per-customer fine-tuning. The bet Engram is making, with $98M and a 13-person team, is that the shift has already started.