Agent Skills in production: wake, search, save as a repeatable workflow
SKILL.md patterns for Cursor and Claude Code agents—session bootstrap, topic recall, atom writes, and guardrails that keep memory consistent across engineers.
Skills are the contract
MCP exposes tools; Skills tell the agent when and how to use them. Without a Skill, models treat memory_search like any other tool—called sporadically, often after context is already wrong.
Engra publishes Skills documented at /docs/memory/skills. Install them in your IDE so every engineer shares the same memory playbook.
Recommended session playbook
1. Session start — call memory_wake_up; inject L0+L1 into working context. 2. User switches topic — memory_recall for the new Topic cluster. 3. Narrow factual question — memory_search with explicit query text. 4. Durable user statement — memory_save_atom with memoryLibraryId, Scope, Topic. 5. Session end — optional save_atom for decisions made during the session.
Guardrails that matter
Never save secrets or raw credentials into atoms—store references and let the agent fetch from your vault at runtime.
Require Topic metadata on every save. Atoms without Topics become orphan embeddings that pollute L3 search.
Log atom IDs returned from wake_up/recall/search in your agent telemetry. When users report wrong answers, you can replay which memory was injected.
Team rollout
Pin the Skill version in your repo. When Engra updates tool schemas, bump the Skill and re-run a short acceptance chat: list libraries, wake up, search, save, list atoms.
Pair Skills with Memory MCP setup for Cursor, or REST for CI agents that cannot run MCP.