A Claude Agent Skill that supplies standing rules for using long-term memory: when to recall before acting, and when to save afterwards.
Connecting a memory tool does not make an agent use it. The tool registers, the session runs, and nothing gets recalled or saved. This skill is the missing part.
- Recall before acting. Which situations warrant a memory lookup before doing the work, and which do not, so recall is not spent on self-contained questions.
- Save after deciding. Which events produce a durable memory: a decision, a correction from the user, a failed approach, a stated preference, an environment fact learned the hard way.
- Write it so it survives. What each entry needs to still be useful weeks later: what, why, when, and where it came from.
- Close the past, do not overwrite it. A superseded memory keeps its validity window instead of being deleted, so old work stays explainable.
- Keep contradictions visible. Two entries that disagree get surfaced with their dates rather than silently resolved.
- Evidence and policy are different weights. One observation is not a rule.
Everything works the same whether memory is:
- a
memory/folder of Markdown notes, one fact per file - a local MCP memory server
- a hosted memory service over MCP
The skill installs no backend and requires no account.
Copy skills/agent-memory-discipline/ into your skills directory:
# project scope
mkdir -p .claude/skills
cp -r skills/agent-memory-discipline .claude/skills/
# or personal scope
mkdir -p ~/.claude/skills
cp -r skills/agent-memory-discipline ~/.claude/skills/As a plugin marketplace:
/plugin marketplace add mnemoverse/agent-memory-discipline
/plugin install agent-memory-discipline@agent-memory-discipline
CC0-1.0. Public domain dedication. Copy it, fork it, rewrite it, ship it in your own collection without attribution.
Written and maintained by the team behind Mnemoverse, which is one hosted memory implementation. The rules are deliberately backend-neutral and were written to be useful without it.