Large context windows are useful, but they are not the same thing as memory.
A larger window gives a model more text to consider in the moment. Memory gives a system continuity. It lets an assistant carry forward decisions, preferences, project state, constraints, and evidence across time. For individual builders, that means less repetition. For enterprise teams, it means AI can become part of a governed workflow rather than a series of disconnected prompt sessions.
The important question is not simply how much context a model can hold. It is who controls that context, how it is inspected, how it is searched, how it is corrected, and how it moves between tools.
That is the thinking behind projects like llm-kosh: local-first memory that stays close to the user, can be searched, can be inspected, and does not depend entirely on a single vendor interface. If AI assistants are going to become daily collaborators, their memory layer should be portable and understandable.
For example, a delivery assistant may remember that a release was approved. If that approval is later withdrawn, retrieving the old note without its status could lead to the wrong action. A useful memory record therefore needs a source, timestamp and lifecycle status, with a clear route for correction or revocation.
Persistent memory also creates costs and risks: stale evidence, private data retained too long, and irrelevant facts added to every prompt. The design should specify what is stored, who can read it, how it expires and how retrieval is tested. A bigger memory is useful only when the system can select the right evidence.