Tools & Products

Tencent Open-Sources Hy Memory 2.0 and MemHub: Cross-Application Context Synchronization for 10+ AI Tools

Tencent released Hy Memory 2.0, an open-source neural memory framework paired with the MemHub desktop app. The system synchronizes persistent user memory across ChatGPT, Claude, Cursor, and local IDEs.

By Maya Lin-Vasquez · 2026-10-18 · 8 min read
Tencent Hy Memory 2.0 and MemHub system architecture graphic demonstrating universal context synchronization across AI clients
Tencent Hy Memory 2.0 and MemHub system architecture graphic demonstrating universal context synchronization across AI clients

On October 18, 2026, Tencent open-sourced Hy Memory 2.0 (derived from its Hunyuan research ecosystem) alongside MemHub, a native desktop application designed to solve persistent context fragmentation across independent AI tools.

Today, developers frequently configure custom instructions inside ChatGPT, repeat coding rules inside Cursor, re-explain database schemas to Claude, and maintain separate prompt files for local terminal tools. Hy Memory 2.0 replaces this repetitive workflow with a single, locally hosted memory kernel that injects synchronized context into any connected client.

Universal Synchronization Architecture

MemHub acts as a local daemon process listening on an encrypted loopback socket. When any supported AI interface initializes a session, the MemHub client injects relevant memories via Model Context Protocol (MCP) and browser extension bridges:

MemHub Local Memory Bridge
┌─────────────────────────────────────────────────────────────┐
│ Developer Workspace (Local Machine)                         │
├──────────────────────────────┬──────────────────────────────┤
│ Supported AI Interfaces      │ MemHub Universal Memory Core │
│ ├── ChatGPT Web / App        │ ├── Local Vector Index       │
│ ├── Claude Web / Desktop     │ ├── Knowledge Graph SQLite   │
│ ├── Cursor / VS Code Copilot │ └── Temporal Decay Pipeline  │
│ └── Local Ollama / vLLM      │                              │
└──────────────────────────────┴──────────────────────────────┘

The framework coordinates across more than 10 widely used tools:

Category Supported Applications Integration Method
Cloud Web Clients ChatGPT, Claude, Perplexity, Gemini Chrome / Firefox Extension Bridge
Code Editors Cursor, Zed, VS Code, JetBrains Model Context Protocol (MCP) Server
Terminal & Local AI Ollama, OpenCode, Antigravity CLI Local REST API (localhost:8484)

Hierarchical Memory Pruning and Decay

A central challenge in persistent memory systems is "context pollution"—stale debugging instructions or transient bugs continuing to bias model behavior weeks later. Hy Memory 2.0 implements a three-tier hierarchical decay model:

  1. Working Memory (Ephemera): Stores active function signatures and recent error traces. Automatically expires after 24 hours of inactivity.
  2. Project Memory (Medium Horizon): Tracks framework versions, API conventions, and database table structures. Updates whenever files are modified in git repositories.
  3. Core Persona Memory (Permanent): Encodes permanent engineer preferences, formatting rules, and language proficiencies. Requires explicit user confirmation to modify or delete.

Privacy and Data Ownership

Unlike proprietary cloud memory implementations that store user session traces on remote laboratory servers, Hy Memory 2.0 runs entirely on the user's workstation:

  • Local Storage: All embeddings and relationship graphs reside in a local SQLite file encrypted with SQLCipher using a user-held passphrase.
  • Zero Cloud Relay: Memory queries resolve against a quantized on-device embedding model without sending telemetry to Tencent servers.
  • Granular Redaction: Users can flag sensitive repository paths (e.g., .env, secrets/) for automatic exclusion from memory extraction.

The complete codebase, documentation, and pre-built binaries for macOS, Windows, and Linux are available on GitHub and Hugging Face.