Design user context as graph-native data (replace context/*.md files)

[unforced]

Problem

User context (profile, preferences, current focus, orientation) currently lives in markdown files at parachute/context/*.md. These files:

• Go stale (now.md last updated Jan 2026, two months ago)
• Are not accessible across sandbox containers (filesystem-bound)
• Have no update mechanism beyond manual editing
• Are referenced via @ imports in CLAUDE.md that may not resolve correctly

Proposed

Move user context to the brain graph as Note nodes with note_type: "context". Auto-load context notes into every chat session's system prompt. Provide a write_note tool so the AI can create/update notes conversationally.

Implementation Plan

Step 1: Add write_note to vault_tools.py

New shared handler:

async def write_note(graph, note_type, title, content, date=None) -> dict:

• For note_type: "context" — MERGE on note_type + title (exactly one note per context title)
• For other note_types — CREATE with generated entry_id
• Sets created_by: "agent", status: "active", created_at: now()
• date optional (required for journals, not for context)
• Content capped at 10K chars

Add tool definition to VAULT_TOOLS list.

Files: vault_tools.py

Step 2: Register write_note in both MCP servers

• Direct MCP (mcp_server.py): Import and add handler in handle_tool_call
• HTTP Bridge (mcp_tools.py): Add to _VAULT_TOOL_NAMES, add case in _handle_vault_tool

Files: mcp_server.py, mcp_tools.py

Step 3: Auto-load context notes in orchestrator

In _build_system_prompt(), after PARACHUTE_PROMPT + CLAUDE.md, before container context:

New method _load_context_notes():

• Query: MATCH (n:Note) WHERE n.note_type = 'context' AND n.status = 'active' RETURN n ORDER BY n.title
• Format as: ## User Context\n\n### {title}\n{content}\n for each note
• Cap total at 8K chars

Files: orchestrator.py

Step 4: Update tool guidance

Add write_note to Browse group in tool_guidance.py. Update list_notes description to mention context notes.

Files: tool_guidance.py

Step 5: Update system prompt vault guidance

Add to PARACHUTE_PROMPT Vault and Memory section:

**Updating context:** When the user says "remember" something about themselves
(preferences, location, current focus), use write_note with note_type='context'
to save it. Context notes are automatically loaded into every session.

Files: orchestrator.py

Step 6: Ensure Note schema exists in core

Move ensure_node_table("Note", ...) to brain_chat_store.py so it runs at startup regardless of whether the daily module is loaded. Daily module can still call it too (idempotent).

Files: brain_chat_store.py

Step 7: Migration script

scripts/migrate_context_to_graph.py — reads existing context/*.md files, creates Note nodes:

• profile.md → Note(note_type="context", title="Profile")
• now.md → Note(note_type="context", title="Now")
• memory.md → Note(note_type="context", title="Preferences")
• orientation.md → Note(note_type="context", title="Orientation")

Skips identity.md and tools.md (redundant with PARACHUTE_PROMPT).

Files: new scripts/migrate_context_to_graph.py

Step 8: Tests

• test_chat_memory.py: Add write_note tests (create, update/merge, size cap, validation)
• test_orchestrator_phases.py: Add context note loading test
• test_tool_guidance.py: Verify write_note in guidance

Not in scope

• Per-session context selection (load all context notes for now)
• Agent-driven context updates (works automatically once write_note exists)
• UI for managing context notes
• Removing old context/*.md files (leave as fallback)

Context

Part of the system prompt overhaul. The CLAUDE.md file has been simplified to no longer reference context files. This issue tracks building the replacement.