AI agent memory should be designed to forget

The firsthand lesson from building Meridian is that a long-term memory system must choose what to forget instead of trying to recall everything. A normal hour of software work can create hundreds or thousands of small events, including editor focus changes, terminal commands, browser moves, documentation reads, log checks, test runs, and window switches.

Saving all of them does not produce useful memory; it creates another context window and moves the overload from the prompt into a database. A better memory pipeline compresses raw activity into related evidence, then into coherent work sessions and engineering decisions, and finally into long-term memory.

Each step intentionally removes detail. Routine traces such as every opened file, command, browser tab, or app switch should be dropped, while rejected hypotheses, reasons behind design choices, production problems that changed the implementation, and investigations that explain the current code should remain.