[FEATURE] Enable skill-scoped auto-memory via skill front-matter

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Learned lessons and insights uncovered while using a skill may end up scattered across multiple projects, bloat user-scoped memory, or otherwise be forgotten due to not belonging in either of those places.

Users must consciously remember to instruct Claude to reflect and improve skills, or must bloat their other CLAUDE.md or SKILL.md files with similar, duplicated instructions.

Proposed Solution

Enable a Claude to write to skill-scoped MEMORY.md files when enabled via the skill's front-matter. Ideally, conditionally insert within existing auto-memory related system prompts, so the same instructions may apply to multiple skills without increasing token consumption for every skill.

Keep it simple, just enabled or not. But, provided it would be effective and safe to do so, reinforce via memory-related portion of Claude system prompt the ability to customize memory maintenance rules for a specific skills via regular instructions in its SKILL.md.

Alternative Solutions

https://github.com/anthropics/claude-code/issues/25469 sounds like it has similar goals, but doesn't aim to leverage the existing auto-memory feature and the context budget already spent on it.

Priority

Low - Nice to have

Feature Category

File operations

Use Case Example

1. While working on a problem using knowledge or processes described by a skill, an unexpected gotcha is encountered and noted in the skill-scoped MEMORY.md
2. Future sessions which invoke the skill also have its MEMORY.md loaded if the skill has enabled it.
3. Performance for the same task will improve and use less tokens, having learned from a mistake previously.
4. Users may later reflect on accumulated MEMORY.md for a skill to consciously decide whether/how to incorporate it into the main skill's files, or improve related scripts or APIs used to avoid the mistake in future (positive feedback loop for either skill or the other scripts, tools or processes it uses).

Additional Context

_No response_