ChatLab: Local AI Archaeology for Your Social Memory

Privacy-First Electron Architecture

ChatLab employs a zero-trust local architecture where sensitive chat data never leaves the main process. The app uses Electron's contextIsolation with a strict IPC bridge for AI operations, ensuring renderer processes cannot access raw message content directly.

Core Abstractions

• ChatArchive: Normalized timeline abstraction decoupled from source format
• MemoryAgent: State machine that reconstructs 'relationship epochs' through RAG
• PrivacyBoundary: Encryption at rest for parsed data with ephemeral processing

Design Trade-offs

The choice of SQLite over browser IndexedDB sacrifices some sandbox security for query performance on multi-GB chat histories—a necessary compromise for handling decade-long WeChat exports that can exceed 10GB of media and text.

The killer insight isn't analyzing chats—it's using agentic narrative reconstruction to turn timestamped logs into episodic memory, answering 'What was my relationship with X during Q2 2023?' rather than just counting message frequency.

Specific Technical Innovations

1. Temporal RAG Pipelining: Implements time-weighted retrieval that prioritizes recent context while maintaining long-term relationship baseline vectors, solving the 'recency bias' in local LLMs with limited context windows (typically 4k-8k tokens).
2. Multi-Modal Local Parsing: WASM-based parsers handle encrypted WeChat SQLite databases and iOS backup formats client-side, eliminating the need to upload sensitive database files to web services.
3. Social Graph Embeddings: Generates dynamic force-directed networks where edge weights represent emotional valence (derived from local sentiment analysis) rather than just message volume, revealing relationship health over time.
4. Differential Privacy Injection: Optional noise addition to timestamp and frequency metadata when users want to share insights (not raw data) with researchers, using ε-differential privacy algorithms implemented in pure TypeScript.
5. Agentic Memory Summarization: Uses a two-stage agent: first a 'librarian' agent extracts relevant message threads, then a 'biographer' agent synthesizes these into coherent relationship narratives with specific quoted evidence.

Local-First Constraints & Benchmarks

Performance is dictated by consumer hardware limits rather than cloud quotas. On a MacBook Pro M3 (18GB RAM):

Scalability Limits

• Memory Ceiling: Vector embeddings for 100k+ messages require ~2GB RAM, pushing the limits of Electron's default heap (4GB) when combined with the renderer process.
• LLM Inference Latency: Running Mistral 7B locally produces analysis 4x slower than API calls but maintains perfect privacy—an acceptable trade-off for the target demographic.
• Parser Blocking: Large media exports (>10GB) currently block the main thread during SQLite decryption; Web Workers are partially implemented but not yet for the cryptographic heavy lifting.

Competitive Landscape

Integration Points

ChatLab functions as a meta-layer rather than a standalone silo:

• Ollama/LM Studio: Native integration for local model management with automatic model pulling (Llama 3, Mistral, Qwen)
• Obsidian: Export relationship reports as markdown with embedded social graphs for personal knowledge management
• Data Freedom: Supports GDPR-style exports from WhatsApp, WeChat, Telegram, iMessage, and Signal (via decryption keys)

Adoption Signals

The 1,302 forks (22.7% fork ratio) is the critical metric here—this isn't passive stargazing. Developers are actively customizing parsers for niche platforms (corporate Slack workspaces, Discord DMs, dating app exports). The bilingual documentation (English/Chinese) captures both the privacy-conscious Western developer market and the massive WeChat user base seeking to escape platform lock-in.