Memory insights
Read the agent's file-based memory (the current project by default, or all projects on request), then give the user a few one-sentence, non-obvious pieces of feedback — constructive or observational — about themselves. Use when the user says "memory insights", "what do you notice about me", "read my memory and tell me something I don't know", or "/memory-insights".From its SKILL.md
npx -y skills add zcaceres/skills --skill memory-insightsAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 7 stars7 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
2.1 KB, as published. Nobody here has run it
memory-insights
The prompt behind this skill: read everything the agent remembers about the user, then hand back a few one-sentence pieces of feedback — constructive or observational — that aren't obvious to them.
How
-
Gather the corpus:
skills/memory-insights/scripts/gather-memories.sh # this project (default) skills/memory-insights/scripts/gather-memories.sh --all # every project + globalIt dumps full memory bodies into one stream. Read the whole thing. Default is the current project's store; reach for
--allwhen the user wants a cross-project read or the current project is thin. If auto-detect misses the project, pass the memory dir with--dir <path>(you know it from context). -
Read through the memories to the person. The store is mostly operational — corrections, project notes, references — but you're not summarizing the store. You're inferring the human behind it: their temperament, their values, their instincts, their blind spots. What would a perceptive colleague notice after reading all of this that the user can't see about themselves?
-
Give 3–5 insights. Each one sentence, non-obvious, constructive or observational. Lead with them, no preamble. Then stop.
Keep it honest
- Read to the person, not the store. If a finding is about memory hygiene or the git workflow rather than the human, cut it.
- Don't flatter, don't fabricate. Every insight traces to something real in the corpus, even when you don't cite it.
- Non-obvious is the bar. If it would make them nod and move on, drop it. Keep the ones that make them pause.