In June, a thirteen-person startup called Engram came out of stealth with $98 million in funding and a valuation of about $600 million. The investor list (General Catalyst, Kleiner Perkins, Sequoia) would be notable on its own. More notable was the name tucked among the angels: Andrej Karpathy. Pulse 2
Engram's pitch is that AI agents are being built on a broken foundation. Every time an agent answers a question about your company, it reconstructs what it knows from scratch, rereading documents, re-deriving context, improvising. Engram wants to replace that improvisation with learned memory: training the organization's knowledge into the model ahead of time, claiming up to 100 times fewer tokens per query. The claim is unverified by independent benchmarks. The check sizes suggest investors aren't waiting for verification. Tech Times
The round capped a twelve-month capital formation wave around what the industry is starting to call the company brain. Mem0 raised $24 million in October to build long-term memory infrastructure, and reports its API traffic grew from 35 million to 186 million calls between the first and third quarters of last year. Cognee closed a €7.5 million seed in February for enterprise-grade memory layers and counts Bayer among its customers. Interloom raised $16.5 million in March; Metacognition AI raised A$10 million in September. Glean, the enterprise knowledge incumbent, sits at a $7.2 billion valuation after last year's $150 million round, with revenue past $100 million. YourStory Pulse 2 Tech Startups TechCrunch
The largest players are building, not just funding. Anthropic put memory stores into public beta for its managed agents in April: memory mounted as a filesystem, versioned, audited, with rollback. Rakuten, an early user, reported first-try errors down 97%. OpenAI rebuilt ChatGPT's memory architecture in June; its internal evaluations show factual recall rising from 67.9% to 82.8%. Dev.to Neowin
Underneath the funding announcements sits a diagnosis that has been building for a year. Karpathy, who coined the term “context engineering” in 2025, has described large language models as suffering from “anterograde amnesia”: brilliant within a conversation, incapable of consolidating anything they learn. Walden Yan of Cognition calls context engineering “effectively the number one job of engineers building AI agents.” Anthropic's engineering team warns of “context rot,” the quiet degradation of agent performance as context windows fill with low-signal tokens. Cognition Anthropic
Enterprises have been paying for the underlying disease for decades. McKinsey's research found employees spend 1.8 hours a day searching for information; IDC puts the figure at 2.5 hours, nearly a third of the working day. Panopto found that 42% of institutional knowledge exists with one person and no one else, and that when that person leaves, colleagues retain only 58% of what they knew. Gallup estimates that replacing an employee costs between half and twice their annual salary. Cottrill Research Manufacturing.net Second Talent
The market sizing is noisy: research firms can't agree whether knowledge management software is a $19 billion or $33 billion market this year, but every forecast points the same direction. Data Bridge sees the category reaching $108 billion by 2033. Enterprise search is growing about 9% a year. Vector databases, the plumbing underneath modern memory systems, are roughly tripling to $9 billion by 2030. And ServiceNow's $2.85 billion acquisition of Moveworks, completed last December, remains the category's landmark consolidation. OpenPR GlobeNewswire EveryTicker
Skepticism is warranted, and the serious kind is specific. Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027, warning of widespread “agent washing.” A sharper engineering critique argues that most so-called memory failures are really documentation failures: that bolting retrieval onto a messy organization adds latency and cost without fixing anything. Gartner Dev.to
Both the hype and the critique point at the same conclusion. Agents that can't remember can't compound: every session starts from zero, every answer is improvised. But memory that only accumulates is a more expensive junk drawer. The companies that win this category won't be the ones that store the most context. They'll be the ones that keep the right context: current, permission-aware, traceable to its source. A company brain isn't valuable because it remembers everything. It's valuable because it keeps what matters, with its reasons.