I've been tinkering with building something similar for myself - though for a generic chatbot, rather than for Claude (not every task is coding, and I'd like to keep !). From other comments (e.g. https://news.ycombinator.com/item?id=46428368, https://news.ycombinator.com/item?id=46427950) suggest that many others are already ahead of me. Any recs for tools, libraries, or approaches that I should learn from or adopt? In particular, I've found that - no matter how direct and clear the system prompt is - models have a tendency to respond verbally as if they've made a tool-call recording some gained-knowledge ("thanks! I'll remember that"), but to not actually return the JSON required to trigger the call by the tool.

Since you've already thought about this problem, I'd love to hear your feedback after giving this skill a try. It should speed up at least your basic need of having to trigger the LLM to store the memory. One of our colleagues has found success asking at the end of a research session what he missed, how he could improve, etc.