I built SAGE because I was running a 12-agent AI pipeline for a cybersecurity CTF platform, and every agent was basically Leonard Shelby from Memento — waking up fresh every session with zero memory of what happened before. I started by building plumbing — consensus-validated memory storage. Infrastructure. But then something unexpected happened. I gave 11 agents 3-line prompts with zero domain expertise, connected them to SAGE, and ran them through sequential tasks. They matched agents with 200-line expert-crafted prompts on quality. But unlike the expert-prompted agents, the memory agents actually got better over time. By run 14, they independently invented a defense technique nobody told them about. The research is across 4 papers (all open access with DOIs in the GitHub repo). The short version: persistent institutional memory enables AI agents to learn from experience without any fine-tuning or weight changes. Same model, same weights — completely different trajectory. SAGE runs as a single binary on your machine. No cloud, no Docker, everything encrypted. Works with any AI that speaks MCP or REST. Full write-up: https://medium.com/@dhillon.andr... Happy to answer any questions about the architecture or the research!

(S)AGE

Persistent encrypted memory for AI that runs locally

Persistent encrypted memory for AI that runs locally