One Board Solves the Session Silos of AI Agents

When an AI coding assistant does work, the most annoying part often isn't that the model isn't smart enough—it's that it doesn't remember where the previous round left off. Code that was half-edited in the last session is completely forgotten in a new session. my-pi-agent 0.19.0 aims to solve this problem, and the approach is simple: put the memory of all sessions on the same board.

my-pi-agent is an open-source AI coding assistant used to change code, run builds, and execute tests in a project. Installation is simple:

pi install npm:@gavin-hu/my-pi-agent

This version has no fancy tricks. The core idea is just one sentence: one board, every session. The agents in all sessions see the same shared board. In the screenshot, the agent is changing return 'Hello, ${name}!'; in two lines of TypeScript to return 'Hello, ${name}, welcome back!';. You can't tell much just by looking at the code, but real task status is attached below the editor: both background jobs have finished, j1 build exited with code 0 and passed; j2 test exited with code 1 and failed. Code changes, build results, and test failure information are all laid out on one board, with project goals, to-dos, and progress statistics below.

Today, as context windows keep getting larger, this direction is actually a bit counterintuitive. What agents lack may not be a longer context. Remembering the global state across sessions is more practical than how many tokens you can cram in at once. While background tasks are running, another session can keep editing code; if tests fail, the board marks them red directly, so you don't have to dig through logs back and forth. For people orchestrating agents across multiple sessions at the same time, this is effectively lifting memory from inside a single session to above all sessions.

If you want to see this feature for yourself, install it, open two sessions on the same project at once, and you'll immediately understand why one board is useful.