agent-memory

Installation
SKILL.md

Agent Memory

Keep one authoritative home for a durable rule, and make clear how it will be read next time. Saving a card does not automatically make an Agent use it.

Choose the action

  • promote: evaluate a lesson and propose a destination and exact wording. Never write memory during promote.
  • apply: write a rule after explicit user confirmation of its content, scope, and exact target.
  • recall: retrieve relevant rules and explain which apply.
  • review: inspect selected memory for stale, duplicate, conflicting, or unsupported rules.

Interpret natural requests by their intent. If the user already explicitly approves exact wording and an exact target in the current conversation, proceed to apply after the checks below; a previous proposal turn is not required. “Remember everything” or “save this globally” without a clear rule and destination is not exact approval. Requests merely discussing or improving this Skill do not authorize memory writes.

Eligibility and authority

Distinguish two inputs:

  • Inferred lesson: require evidence, reusability, an observable action, stable scope, and safe wording. Prefer two independent occurrences; one deterministic, high-impact failure can justify a narrow rule. State the threshold used. Repeated discussion of the same event is one occurrence.
  • Explicit durable user rule or preference: the user's statement is authority for the preference, not proof of an empirical claim. It need not recur twice. Retain its scope; do not generalize a project choice to all projects.
Installs
1
GitHub Stars
1
First Seen
2 days ago
agent-memory — chrichuang218/agent-handoff