On June 23, 2026, an eight-month-old startup named Engram emerged from stealth with $98 million in Series A funding and a $600 million valuation. Kleiner Perkins led the round, joined by General Catalyst, Sequoia Capital, Factory, Modern, Amplify Partners, and Neo, plus notable angels including Andrej Karpathy (OpenAI co-founder, recently joined Anthropic) and Assaf Rappaport (CEO of Wiz). The thesis is straightforward: every enterprise’s AI is a brilliant stranger that re-reads the same documents and re-learns the same context on every query, and that waste has become one of the largest line items in enterprise technology budgets.
This round is the latest signal in a pattern Enera has been tracking closely: institutional capital is flowing toward tools that solve the operational problems of enterprise AI, not just the capability gaps. After covering the agentic GTM funding wave of June 2026 and the widening gap between AI agent adoption and enterprise readiness, Engram adds a third dimension: the infrastructure cost of running AI agents at scale.
The problem: AI that forgets everything, every time
Enterprise AI today operates on what Engram CEO Dan Biderman calls the “genius stranger model.” A frontier model is extraordinarily capable in the abstract, but it knows nothing specific about your organization. When your legal team’s AI agent reviews a 70,000-word contract, the model’s internal representation of that document can balloon to over 100 gigabytes of internal state, approximately 250,000 times the size of the original file, according to Engram CTO Sabri Eyuboglu in the company’s launch statement. The model regenerates that state from scratch every time a query arrives.
Multiply that by thousands of employees running agents across legal, sales, operations, and marketing, and the arithmetic becomes unsustainable. CNBC’s Samantha Subin, reporting on the round, noted that enterprises are “starting to crack down on untamed AI usage by developers,” with token bills surfacing as a material budget problem. More sophisticated models require more tokens per query, and many enterprises discovered in early 2026 that costs scaled faster than the value returned.
Biderman explained the timing in interviews with Calcalist Tech: “When we started building the company, token costs were not a major concern for most people in the industry. But at the beginning of this year, with the release of new capabilities in models such as Claude and others, token consumption increased dramatically. Suddenly, cost became a critical issue for organizations.”
How Engram works: learned memory, not retrieval
Engram’s approach is architecturally distinct from retrieval-augmented generation (RAG) and long-context window stuffing. Rather than fetching documents at query time, Engram pre-trains models on an organization’s specific world: its documents, workflows, terminology, and operating patterns. The result is a compact, reusable memory embedded in the model itself, built once and reused on every subsequent query.
Biderman described the ambition in Engram’s launch press release: “Whatever the AI knows about you is improvised on the spot, a sticky note about your past, a document pulled mid-conversation. If we can anticipate your interactions, we can prepare memories ahead of time instead of pasting them on the fly.”
The company claims its models match or outperform frontier models while using 1 to 10 percent of the tokens, a reduction of 10x to 100x. Kleiner Perkins partner Leigh Marie Braswell described it in the firm’s investment perspective: “Engram comes in and basically maps out your organization and offers orders of magnitude cheaper output.”
A compounding property makes this more than a cost play. The longer an organization uses Engram, the more specialized and proprietary the model’s memory becomes. Generic frontier models reset on every conversation; Engram models improve with use and cannot be extracted by the model provider. As Biderman put it: “Today, if you wanted to make your AI better, there is almost nothing you can do. Your AI gets better when the model behind it gets better. How you use it has almost no effect. We are building toward a different future.”
The founding team: research depth as a moat
Engram was founded in October 2025, coming directly out of Stanford University’s AI lab. Biderman completed postdoctoral research at Stanford under Chris Re, one of the most influential figures in modern machine learning and an Engram co-founder. His co-founders bring complementary depth across the specific research problems the company is attacking.
CTO Sabri Eyuboglu (Stanford PhD under Chris Re) developed “Cartridges,” a method for compressing large document collections into small, reusable model memories. Jessy Lin (Berkeley PhD, former researcher at Meta’s FAIR lab) created “Active Reading,” a training approach that teaches models to study material deeply rather than simply retrieve and re-process it. Jack Morris (Cornell PhD, also from FAIR) specializes in retrieval and LLM memorization. Scott Linderman is a tenured Stanford professor of statistics and neuroscience, with foundational work on state space models as an efficient alternative to transformers for long-sequence tasks.
Calcalist Tech reported that the company employs just 13 people at its $600 million valuation, a ratio that reflects both the capital efficiency now possible for AI-native companies and the deliberate scarcity of researchers with the specific expertise Engram requires. Biderman told Calcalist: “The talent we are looking for is extremely rare. The people we have recruited are among the best in the industry.”
Enterprise partnerships at launch
Engram launched with three anchor commercial partners already in production tests.
Microsoft is testing Engram’s models inside Microsoft 365, exploring how a learned memory layer could bring organizational knowledge into tools used by hundreds of millions of knowledge workers. The partnership includes dedicated GPU capacity across Microsoft’s Dapple and Azure infrastructure. Jason Graefe, Corporate Vice President at Microsoft AI Partner Catalyst, described the goal: building AI efficient enough to “power the long-running, proactive agents we believe every knowledge worker will eventually rely on.”
Notion is running Engram inside custom agents for enterprise workspaces. Co-founder Simon Last cited the token cost problem directly: “Our enterprise customers are running long-lived agents across their Notion workspaces, and that kind of always-on work can burn through tokens fast, even for something as simple as triaging a task. We are already seeing them approach frontier quality while using an order of magnitude fewer tokens.”
Harvey, the legal AI company, is building what co-founder Gabe Pereyra described as “learned enterprise memories that are secure, cost-efficient, and turn unstructured context into compounding agentic knowledge bases.”
What enterprise AI leaders should act on
The Engram round carries three practical implications for enterprise teams planning AI deployments in the second half of 2026.
Token cost is now a strategic design input. When you deploy AI agents across sales, legal, operations, and marketing simultaneously, the token bill compounds at every step. Enterprises that evaluated AI tools in 2024 and 2025 on accuracy and capability alone are now finding that operational cost reshapes the ROI calculation significantly. Any organization building an AI-native operating model needs memory architecture in the design from the start.
Proprietary organizational memory is an emerging competitive advantage. Engram’s models do not just reduce costs: they accumulate intelligence that is unique to each organization and cannot be extracted by the model provider. Fine-tuning and prompt engineering modify how a model behaves, but they do not create the kind of compounding organizational memory Engram is building. The more an organization uses it, the further ahead it gets.
The infrastructure layer of agentic AI is still taking shape. The agentic GTM platforms funded in June 2026 represent the application layer: tools that take action in existing stacks. Engram is building the memory substrate those agents need to function at enterprise cost points. The two layers are complementary, and the next 18 months will likely see both mature simultaneously.
For enterprise teams auditing their AI spend today, the immediate question is concrete: what share of your current token costs comes from re-reading context your agents have already processed? If you do not know the answer, that is itself the signal.
If your organization is designing AI systems for scale, the team at Enera works on exactly this layer: cost-aware architecture, operational intelligence, and AI systems built for enterprise reality.