Engram raises $98M to build memory layer for AI agents

Engram is betting that the missing piece of enterprise AI isn’t another foundation model but a durable memory layer that sits between models and business data. That’s strategically important in the race to AGI because long-horizon, multi-session reasoning is one of the core weaknesses of today’s LLM-centric systems. By training models to pre-compress institutional knowledge into a compact memory store, Engram is attacking the cost and latency penalties of repeatedly re-ingesting the same context.

This funding round effectively validates “agent memory infrastructure” as its own category alongside vector databases and retrieval systems. The investor roster—top-tier US venture firms plus prominent AI figures—suggests confidence that memory will be a defensible layer even as base models commoditize. Partnerships with Microsoft 365, Notion and legal AI startup Harvey also indicate that Engram wants to become the standard way enterprise agents remember and forget over time.

If Engram and similar systems succeed, they’ll make agents cheaper, faster and more reliable on proprietary data, shifting some value away from closed frontier models toward orchestration and memory. That doesn’t change what frontier models can do in principle, but it makes near-frontier systems much more competitive for high-value enterprise work, which in turn can redirect capital and experimentation into agent architectures rather than sheer model scale.