Add Memory Upload Intake and Meta/Facebook data analyzer

[besfeng23]

Goal

Add a safe Memory Upload Intake pipeline that can accept uploaded data, analyze it, organize it into the best-fit memory structure, and propose append-only memory patches for review.

Core principle

Do not blindly dump uploaded files into memory.

Required flow:

Upload data → extract → analyze → classify → organize → propose memory → human review → append approved memory only

Supported upload sources

• Meta/Facebook data exports
• Messenger conversations
• Facebook posts, comments, reactions, profile/activity data
• PDFs
• screenshots
• emails
• chat histories
• CSV/XLSX files
• contracts and business documents
• project notes
• code docs
• relationship chat histories
• AU/fiction canon files

Meta/Facebook data fit

Meta/Facebook export data is a high-priority fit because it contains:

• long relationship and communication timelines
• Messenger conversations
• people/entity maps
• timestamps and repeated interaction patterns
• reactions, comments, posts, profile events
• relationship signals and social graph clues
• duplicates and noisy JSON/HTML structures that need cleanup

Required Meta/Facebook analysis behavior

When a Meta/Facebook data pack is uploaded, the system should:

1. Detect export format and file structure.
2. Parse Messenger JSON/HTML safely.
3. Normalize timestamps and participants.
4. Separate chats, posts, comments, reactions, profile data, and activity logs.
5. Extract people, events, preferences, relationship patterns, repeated facts, decisions, conflicts, timelines, and unresolved threads.
6. Deduplicate repeated records.
7. Identify sensitive or risky material.
8. Separate real-life facts from AU/fiction content.
9. Score confidence and source every proposed memory.
10. Generate reviewable memory patches.

Namespaces

The upload system must preserve namespace boundaries:

• real_life
• business
• relationships
• projects
• legal
• financial
• technical
• fiction_au
• nsfw_fiction
• temporary_scratch

Meta/Facebook real user data must default to real_life / relationships unless explicitly marked as fictional or AU.

Review requirements

The system must show the operator:

• what was found
• what should become memory
• what should remain file-only reference
• what is uncertain
• what is sensitive
• what conflicts with existing memory
• what should be rejected

Only approved patches may be appended.

Safety requirements

• No automatic blind save.
• No cross-contamination between real-life memory and AU/fiction memory.
• No unsupported claims stored as fact.
• No public ingestion by default.
• All appended memory must include source, confidence, namespace, audit event, and review decision.
• High-risk relationship, legal, financial, sexual, identity, or accusation content requires review.

Future MCP fit

This should later become an MCP-capable upload tool:

memory.upload_data_pack
memory.analyze_upload
memory.propose_patches
memory.review_patch
memory.append_approved_patch

Initial release should remain internal/admin only.