One thing I’d want to preserve carefully is the type of memory, not just its relevance score. In coding-agent workflows, a note can be a durable architecture decision, a temporary debugging hypothesis, a personal preference, a failed path, or a constraint that only applied before a refactor.

Those should probably age and prune very differently.

The failure mode I keep seeing is not only bloated context. It is stale or low-authority context being reused with high confidence. A small memory that knows why something was kept, who it applies to, and when it should stop being trusted seems more useful than a large memory with better retrieval.

In coding-agent work I see a similar pattern. The best outcomes usually happen when the agent is forced to study concrete source material first: real repos, real docs, real examples, and the constraints behind them. The worse outcomes happen when it generates a plausible path from a vague prompt and never has to reconcile that with existing practice.

For learning, I imagine the same thing matters: the LLM should help structure the path and explain the friction, but the learner still needs to touch the code and compare against sources.

The source-backed part feels more important than the generated tutorial part.