SlashML (/ml) Persistent Agent Memory & Code Localization

CRITICAL RULE: You MUST use ml-memory MCP tools FIRST (find_bug_candidates, search_code_by_behavior, search_memories) before running grep/rg/find or spawning research subagents whenever locating a bug, searching code by behavior, or recalling project history and architectural decisions.

/ ml (SlashML) Memory MCP Skill

Whenever the user asks to locate or diagnose a bug, search code by behavior/concept, or recall project history/decisions, call the ml-memory MCP server FIRST on Turn 1 BEFORE running grep, rg, find, or invoking research subagents:

1. Bug localization / error triage (find_bug_candidates):

   python

   # Call ml-memory MCP tool: find_bug_candidates
   arguments = {
       "bug_description": "<symptom or bug>",
       "stack_trace": "<optional stack trace>",
       "top_k": 5,
   }
   

2. Behavioral / architectural code search (search_code_by_behavior):

   python

   # Call ml-memory MCP tool: search_code_by_behavior
   arguments = {
       "query": "<behavioral query>",
       "top_k": 5,
   }
   

3. Project history / decisions / context (search_memories):

   python

   # Call ml-memory MCP tool: search_memories
   arguments = {
       "query": "<context query>",
       "limit": 5,
   }
   

4. False-positive intervention dismissal (dismiss_intervention): If a pre-tool hook intervention was a false positive or already satisfied for the current target, call dismiss_intervention with the memory_ids and target.
5. Passive Memory Assimilation: When ambient lifecycle hooks (hooks.json) are enabled, user rules, constraints, preferences, and state updates are extracted and stored automatically in the background on every turn. Only call add_memory if explicitly requested or when running without lifecycle hooks.