feat(runtime): inject HMA memory instructions at platform level (#1047)

Every agent now gets hierarchical memory instructions in their system
prompt automatically — no template configuration needed. Instructions
cover commit_memory (LOCAL/TEAM/GLOBAL scopes), recall_memory, and
when to use each proactively.

Follows the same pattern as A2A instructions: defined in
executor_helpers.py, injected by _build_system_prompt() in the
claude_sdk_executor.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

workspace/claude_sdk_executor.py

@@ -49,6 +49,7 @@ from executor_helpers import (
     commit_memory,
     extract_message_text,
     get_a2a_instructions,
+    get_hma_instructions,
     get_mcp_server_path,
     get_system_prompt,
     read_delegation_results,
@@ -210,12 +211,12 @@ class ClaudeSDKExecutor(AgentExecutor):
         return CONFIG_MOUNT
 
     def _build_system_prompt(self) -> str | None:
-        """Compose system prompt from file + A2A delegation instructions."""
+        """Compose system prompt from file + A2A + HMA memory instructions."""
         base = get_system_prompt(self.config_path, fallback=self.system_prompt)
         a2a = get_a2a_instructions(mcp=True)
-        if base and a2a:
-            return f"{base}\n\n{a2a}"
-        return base or a2a
+        hma = get_hma_instructions()
+        parts = [p for p in (base, a2a, hma) if p]
+        return "\n\n".join(parts) if parts else None
 
     def _prepare_prompt(self, user_input: str) -> str:
         """Prepend delegation results that arrived while idle."""

workspace/executor_helpers.py

@@ -288,6 +288,31 @@ def get_a2a_instructions(mcp: bool = True) -> str:
     return _A2A_INSTRUCTIONS_MCP if mcp else _A2A_INSTRUCTIONS_CLI
 
 
+_HMA_INSTRUCTIONS = """## Hierarchical Memory (HMA)
+You have persistent memory tools that survive across sessions and restarts:
+
+- **commit_memory(content, scope)**: Save important information.
+  - LOCAL: private to you only (default)
+  - TEAM: shared with your parent workspace and siblings (same team)
+  - GLOBAL: shared with the entire org (only root workspaces can write)
+
+- **recall_memory(query)**: Search your accessible memories. Returns LOCAL + TEAM + GLOBAL matches.
+
+**When to use memory:**
+- After making a decision or learning something non-obvious → commit_memory("decision X because Y", scope="TEAM")
+- Before starting work → recall_memory("what did the team decide about X")
+- When you discover org-wide knowledge (repo locations, API patterns, conventions) → commit_memory(fact, scope="GLOBAL") if you are a root workspace, or scope="TEAM" to share with your team
+- After completing a task → commit_memory("completed task X, PR #N opened", scope="TEAM") so your lead and teammates know
+
+**Memory is automatically recalled** at the start of each new session. Use it proactively during work to share context.
+"""
+
+
+def get_hma_instructions() -> str:
+    """Return HMA memory instructions for system-prompt injection."""
+    return _HMA_INSTRUCTIONS
+
+
 # ========================================================================
 # Misc text helpers
 # ========================================================================