nossa-y/activity-frames

Activity frames gives an AI agent a memory of what you have actually been doing on your computer, not just what you have typed into a chat. Normally an agent starts every conversation blind, with no idea what you worked on in the last few hours. This tool records your screen locally, turns that raw recording into structured summaries of your activity, and hands those summaries to any agent that supports MCP, a common protocol for connecting tools to AI models. Instead of storing thousands of raw screenshots, the tool compiles your activity into what it calls frames: short episodes describing what app you were in, which website you visited, what kind of page it was, how long you spent there, and how much you typed or clicked, all without any guessing about your intentions. That structured episode can then be turned into a short block of plain text describing your day, something like a timeline of which apps and sites you used and for how long, which can be pasted straight into an agent's prompt so it understands your recent context. The project is explicit about honesty in what it reports. Anything derived automatically and deterministically from the raw recording is labeled as measured, while anything involving interpretation or guessing, such as figuring out that a series of page visits meant you were job hunting, is treated as a separate, clearly labeled and optional layer, so facts and guesses never get mixed together silently. Everything happens entirely on your own machine. Nothing is uploaded anywhere, no cloud service is involved, and no language model is used to produce the memory itself, only plain code, although the underlying screen capture tool does use on device text recognition to read what appears on screen. Typed text content is excluded from output by default unless you explicitly turn it on. Setup involves installing the tool through pip, starting the recorder, and either using its command line tools or connecting it to an agent like Claude Code as an MCP server. It currently runs best on Apple Silicon Macs, with less tested support for Intel Macs and Linux.