Memory Architecture

Picture three physical shelves: a journal, a stack of index cards, and a folder of current briefs.

The journal captures what happened, in sequence, without cleanup. The index cards hold specific facts — compact, dated, updateable. The current briefs are synthesized snapshots of what is true right now. These are different things, and mixing them breaks all three.

Journals alone are not enough. They preserve everything, but retrieval gets expensive. To answer a precise question from six weeks ago, you have to scan long narrative history and hope the answer surfaces. That is slow and error-prone at volume.

Index cards solve that. A good atomic fact is small and explicit: "On March 24, the project launch moved from April 3 to April 17." When the date changes again, you do not overwrite the old card. You add a new one and mark the previous version superseded. Now the system knows both the current state and the full change history. Auditability without noise.

Current briefs solve the human problem: nobody wants to read raw index cards as a daily habit. A synthesis pass turns active facts into a clean snapshot — what is true now, legible without archaeology.

When teams collapse these layers, memory degrades fast. Everything in daily notes means broad recall, poor precision. Vague or undated facts mean nobody can trust which version is current. Overloaded summaries stop being useful as operating references.

The separation is not academic. It is why an agent can recover after a long pause, hold preferences stable across weeks, and avoid relearning the same lessons from scratch. Better structure produces steadier judgment, not just better storage.

Good memory is not perfect recall. It is a system where truth can evolve without losing the thread.