Memory Skill

Use Cortex memory only through MCP tools.

Non-Negotiable Rules

1. Never treat memory as files on disk. Do not use file tools or grep for memory data.
2. Always set store explicitly on memory/category operations.
3. Load memory before decisions: list → narrow category → fetch only relevant entries.
4. Write atomic memories (one fact/decision per memory), update instead of duplicating.
5. Save small temporary trace memories often while working (plan updates, findings, decisions-in-progress) so future sessions can resume quickly.
6. Set expiration for temporary context; leave durable knowledge without expiry.
7. Refuse or redact secrets/sensitive data before storing memory.

Default Operating Loop

1. Discover stores: cortex_list_stores()
2. Explore relevant categories: cortex_list_memories(store: "...", category: "...")
3. Read only what is needed: cortex_get_memory(...)
4. Persist key deltas frequently as small entries: cortex_add_memory(...) or cortex_update_memory(...)
5. Add expiry to temporary trace entries and prune as needed

Prompt-to-Workflow Routing Examples

• "Start this task and load relevant memory first" → references/workflows/session-start.md
• "We just ran cortex init for this repo; create initial memories" → references/workflows/new-project-init.md
• "Capture this architectural decision so we remember it" → references/workflows/capture-knowledge.md
• "From this conversation, save my preferences and project rules" → references/workflows/synthesize-facts.md
• "Clean up memory structure; categories are messy and duplicated" → references/workflows/review-hygiene.md
• "Wrap up: expire temporary notes and prune stale entries" → references/workflows/session-end-cleanup.md

References

• Fundamentals (store choice, hierarchy, naming, loading, expiry, schema): references/fundamentals.md
• MCP operations cheat sheet: references/operations.md
• Workflow index: references/workflows.md