memory-management

Memory Management Skill

Use AutoMem to maintain persistent context across Claude Code sessions.

Tool examples below use short names (recall_memory, store_memory); call them on whichever AutoMem MCP server is wired in (plugin installs namespace them as mcp__plugin_automem_memory__*, user-level servers as mcp__memory__*).

This skill teaches the current AutoMem playbook: Recall early, store durable outcomes, avoid session-summary noise.

Phase 1: Session Start (Recall)

Always Recall For

• Project context questions (architecture, tooling, deployment)
• Architecture discussions or decisions
• User preferences and code style
• Debugging issues (search for similar past problems)
• Refactoring (understand why current structure exists)
• Integration or API work (check past implementations)
• Performance optimization discussions

Standard Recall Pattern

• Preferences first: recall with tags: ["preference"], limit: 20, sort: "updated_desc"
• Task context second: one semantic query built from actual nouns in the request, optional project slug if unambiguous
• Debugging on demand: recall with the error text as the semantic query and NO tags — bugfix/solution tagging is incomplete, and a tag gate hides cross-corpus fixes

Skip Recall For

• Pure syntax questions ("How does Array.map work?")
• Trivial edits (typos, formatting, simple renames)
• Direct factual queries about current code
• File content requests that can be answered by reading

Recall Examples

// Preferences first
recall_memory({
  tags: ["preference"],
  limit: 20,
  sort: "updated_desc"
})

// Task context recall
recall_memory({
  query: "authentication timeout PostgreSQL auth.ts retry logic",
  tags: ["myapp"],   // drop if ambiguous
  time_query: "last 90 days",
  limit: 30
})

// Debug similar errors (no tag gate — a hard gate hides cross-corpus fixes)
recall_memory({
  query: "TimeoutError authentication request timed out",
  limit: 20
})