Maybe, every night, you send the AI off to "sleep" where it uses those in cache "memories" to influence the long term weights [1].

[1] https://www.pnas.org/doi/10.1073/pnas.2220275120

Context self-distillation does exist, but as is, it's used mostly in training rather than as a part of a continuous learning mechanism.