My agent had no idea what I did all day. I compiled 55 days of screen capture into episodic memory it can read: 88x fewer tokens, 68ms per day, no LLM.

Check more stories related to machine-learning at: https://hackernoon.com/c/machine-learning.

You can also check exclusive content about #ai-agents, #open-source, #python, #artificial-intelligence, #software-development, #automation, #screen-activity, #hackernoon-top-story, and more.

This story was written by: @nossaiyamu. Learn more about this writer by checking @nossaiyamu's about page,

and for more stories, please visit hackernoon.com.

AI agents remember conversations but not what their user actually did all day. I compiled 55 days of my own screen capture into activity frames (bounded episodes with apps, pages, durations, input counts) using deterministic code, no LLM in the pipeline. One day of raw capture is 126,812 tokens; the compiled context block is 1,441 (88x smaller), builds in 68ms, and is byte-identical across runs. The schema keeps measured facts and inferred labels in separate tiers so memory stays auditable. Open-source implementation with an MCP server included.