Every Agent Memory Repo Stores What You Know
Every Agent Memory Repo Stores What You Know
Open any agent memory repository on GitHub and you will find the same pattern: the user tells the agent things, the agent stores them, the agent retrieves them later. This is note-taking with extra steps.
The Missing Data Source
The most valuable information about how you work is not what you say - it is what you do. Your browser history shows which documentation pages you visit repeatedly. Your file access patterns reveal which parts of the codebase matter. Your application switching patterns expose your real workflow, not the one you think you have.
This data already exists. Your browser stores history. Your OS tracks file access. Your editor logs which files you open. But agent memory systems ignore all of it because they are designed around explicit knowledge capture instead of implicit behavior observation.
Browser Data Into SQLite
The practical implementation is straightforward. Extract browser history, bookmarks, and frequently visited pages into a SQLite database. Rank entries by access frequency, not recency. A page you visit every day for a month is more important than a page you visited once yesterday.
SQLite is the right store because it is local, fast, and requires no server. Your agent queries it directly. No vector database. No embedding pipeline. Just SQL queries against structured data that already exists on your machine.
Frequency Over Recency
Most memory systems prioritize recent information. But frequency reveals importance. If you visit the same API documentation page fifty times, that page is critical to your work. Your agent should know it exists and surface it proactively.
Ranking by access frequency transforms agent memory from a recall system into an anticipation system. The agent does not wait for you to ask. It already knows what you need.
Fazm is an open source macOS AI agent. Open source on GitHub.