Memory in AI systems is not a technical problem. It is an editorial one. What these systems retain, discard, or emphasize will shape what we collectively know — and forget.
The people designing memory architectures for large language models are making choices that would, in any other context, be called editorial decisions. They are deciding what counts as signal, what decays, and whose experiences are represented in the foundation that everyone else builds on.
The early internet had a similar moment. The decisions made about search indexing — what got surfaced, how freshness was weighted, whose pages got crawled — had consequences that took two decades to fully understand.
We are at a comparable inflection point, with less time to figure it out.