Engram

About Engram

Engram is a San Francisco-based AI research company building AI models that learn from a user's context and deeply understand their work. Its argument is that today's models know mostly what was on the public internet and re-read the same company documents on every query, so they become more expensive and more confused as context grows and forget everything when a chat closes. Engram's core product is a learned memory layer that separates organisational knowledge from model reasoning so that enterprise AI systems do not have to re-process full context each time.

Instead of spending huge training compute on public data, the company starts from strong pre-trained models and spends training compute on the context a customer cares about, so each model effectively studies documents, workflows and institutional knowledge in depth and draws connections that had not been made. It says this lets models skip re-gathering context and be many times more token-efficient, behaving like a teammate who already knows the work. The team trains and tests its models on its own GitHub, Slack and Notion data and works with a design partner in the enterprise software ecosystem.

Its north star is a single training algorithm that can continually absorb arbitrary amounts of data into a model, with company data currently processed daily and a goal of retraining hourly and eventually every minute, which it acknowledges is still an open problem at scale. LinkedIn describes the company as scaling compute on your context, with a small team in San Francisco.

Engram sells its memory technology to enterprises and AI platforms that run large language model workloads and want lower inference cost and better personalisation, positioning itself in the AI infrastructure and inference optimisation market as an alternative to ever-larger context windows and repeated retrieval.