Bootstrap: conversational onboarding + system discovery for day-zero memory

[kovtcharov]

Summary

Solve the cold-start problem. When the agent has no memory (first launch), bootstrap it through a guided conversation and optional system discovery so it's useful from the first interaction.

Spec: docs/spec/agent-memory-architecture.md → "Bootstrap: Day-Zero Onboarding"

Depends on: #542 (MemoryStore), #543 (MemoryMixin)

Phase 1: Conversational Onboarding

Guided question flow that runs on first launch or via gaia memory bootstrap:

• Name, role, timezone → fact (global)
• Use cases → suggest contexts (work, personal, etc.)
• Communication preferences → preference (global)
• Tools/languages → fact (work context, with entity linking)
• Adaptive follow-ups based on answers (student vs engineer vs designer)

Implementation: run_bootstrap_conversation() method on MemoryMixin.

Phase 2: System Discovery

File: src/gaia/agents/base/discovery.py

Opt-in system scanning with user review before storing:

Source	What it reads	Sensitive?
File system	Folder names + extensions (not contents)	No
Git repos	.git/config remotes, branch names, languages	Partial
Installed apps	Windows registry, Start Menu shortcuts	No
Browser bookmarks	Chrome/Edge/Firefox bookmark files	Partial
Browser history	Last 30 days URLs (not page content)	Yes — all auto-flagged
Email accounts	Credential store addresses only	Yes

All discoveries returned as dicts with source='discovery', confidence=0.4.
User reviews and approves/rejects each item before storage.

Auto-classification into contexts via heuristics (~/Work/ → work, social media → personal).

CLI

gaia memory bootstrap              # Full bootstrap
gaia memory bootstrap --chat-only  # Conversation only
gaia memory bootstrap --discover   # Discovery only (re-scannable)
gaia memory bootstrap --reset      # Clear source='discovery' items (with confirmation)
gaia memory status                 # Memory stats

Privacy Safeguards

• No file contents, no email content, no browser page content
• Everything local — nothing transmitted
• Auto-flag sensitive discoveries
• User review required before any storage
• All items tagged source='discovery' for bulk delete
• --reset only deletes discovery items, preserves user-edited items

Acceptance Criteria

• Conversational onboarding asks adaptive questions
• SystemDiscovery scans 6 sources on Windows
• Discoveries presented for user review before storing
• Sensitive items auto-flagged
• Context auto-classification works
• CLI commands work (bootstrap, status)
• Re-run doesn't duplicate (dedup handles it)
• --reset is safe (only source='discovery', with confirmation)