As corporate America tightens its grip on runaway AI spending, a startup barely eight months old is betting there is a fortune to be made in helping companies do more with fewer tokens. Engram, which describes itself as the "learned memory" of AI, announced on June 23, 2026 that it has raised $98 million from a marquee list of investors.
The round drew backing from General Catalyst, Kleiner Perkins, and Sequoia, as well as OpenAI co-founder Andrej Karpathy, who recently joined rival lab Anthropic. The funding signals that top-tier venture capital still sees room for a new player in an increasingly crowded field — provided it can solve a problem the industry's biggest names have not: the rising cost of running AI models. Per maggiori informazioni su questa storia, consulta la nostra tendenze IA.
Gli obiettivi dell'engramma del problema dei costi
New and more sophisticated AI models are proving pricier than the generations that preceded them, challenging the once-dominant assumption that greater scale would inevitably drive costs down. Companies that eagerly adopted AI tools are now confronting the bill, with some reining in usage by developers as token consumption strains budgets.
Engram's pitch is that its models can recall organization-specific workflows and context to anticipate questions and deliver smarter responses at a fraction of the cost. The company claims its models can match or outperform frontier labs using up to 100 times fewer tokens, the computational currency that determines how much each AI query costs to run.
Leigh Marie Braswell, partner di Kleiner Perkins, ha delineato l'opportunità senza mezzi termini. "C'è questa esplosione di dati, un'esplosione di costi", ha detto, secondo la CNBC. "Engram arriva e fondamentalmente mappa la tua organizzazione e offre risultati di ordini di grandezza più economici."
Primi successi con grandi nomi
Despite having been founded less than a year ago, the 13-person company has already built a client roster that would be the envy of far larger rivals. Tra i clienti figurano Microsoft, Notion e Harvey, la startup legale di intelligenza artificiale. That early adoption by enterprises — and by another AI company in Harvey — suggests Engram's technology is being used not as a novelty but as core infrastructure for real workloads.
La startup prende il nome dalle neuroscienze, dove per “engramma” si intende la traccia fisica che un ricordo lascia nel cervello. The metaphor is apt: Engram's technology is designed to give AI systems a durable memory of an organization's specific knowledge, so they do not have to relearn context from scratch on every interaction.
Un fondatore ossessionato dalla memoria
Il co-fondatore e CEO di Engram, Dan Biderman, ha un legame personale con il problema che la sua azienda sta risolvendo. According to CNBC, Biderman's interest in memory began as a child, when he tried to help his grandmother — who had lost her memory — remember small facts about him and his siblings. Quel fascino permanente per il funzionamento della memoria è ora alla base dell’approccio tecnico dell’azienda volto a rendere l’intelligenza artificiale più economica e più consapevole del contesto.
Biderman è anche una figura riconoscibile nella comunità di ricerca sull’intelligenza artificiale, noto per il suo lavoro sulla comprensione e sul miglioramento di grandi modelli linguistici. That technical credibility has likely helped Engram attract both elite investors and demanding enterprise customers in its first months of operation.
Cosa compra il denaro
Engram said it plans to use the new funding to support compute and talent — the two resources that define a startup's ability to train, refine, and serve AI models. For a company claiming dramatic efficiency gains, the need for substantial compute highlights that building better memory systems still requires serious engineering and infrastructure investment.
The round also reflects a broader trend in AI venture funding: investors are increasingly drawn to picks-and-shovels companies that make existing AI deployments cheaper or more effective, rather than betting exclusively on yet another foundation model. Startups focused on inference efficiency, memory, and cost optimization are commanding attention precisely because the era of unconstrained AI spending appears to be ending.
Il panorama competitivo
Engram entra in un mercato in cui la definizione di "memoria AI" è ancora contestata. Some companies attack the cost problem through better model routing, sending easy queries to cheap models and reserving expensive ones for hard tasks. Altri comprimono il contesto utilizzando tecniche di recupero o memorizzazione nella cache specializzata. Engram's approach — building dedicated memory models that learn an organization's patterns — represents a distinct bet that remembering context efficiently is itself a product category.
Resta da vedere se l'audace affermazione della società di eguagliare la performance di frontiera a un centesimo del costo simbolico reggerà sotto esame. But with Microsoft and other major enterprises already on board, and with investors like Karpathy — whose career spans the founding of both OpenAI and a stint at Tesla — placing their capital behind it, Engram has quickly established itself as a startup to watch in the economics-of-AI race.
---
Rimani Aggiornato sull'IALe ultime notizie, analisi e scoperte sull'intelligenza artificiale — tutto in un posto.
Leggi altre notizie sull'IA →
