jiabaobei/memory-bridge

What it solves

MemoryBridge provides a cross-device, cross-platform shared memory layer for AI agents. It solves the problem of fragmented context when switching between devices (e.g., moving from a mobile phone to a PC), eliminating the need to manually copy-paste chat histories or re-explain backgrounds to an AI. Unlike single-machine memory systems, it ensures that semantic memory follows the user rather than the the application.

How it works

The system implements a Cross-Device Semantic Memory Continuity Architecture (CDSMP). It uses a "channel"—a folder on a cloud drive (like OneDrive or Nutstore)—as a meeting point for devices to exchange encrypted differential packages of memory.

Key technical mechanisms include:

• Deterministic Key Derivation: Devices derive the same encryption key from a hardcoded seed using PBKDF2-HMAC-SHA256, allowing zero-config synchronization.
• Hybrid Retrieval: It combines vector search, keyword matching, and a Semantic Association Network (SAN) graph, fusing results via Reciprocal Rank Fusion (RRF).
• Content Freezing: To prevent hallucinations, the system never rewrites or summarizes original memories; it only adjusts structural parameters.
• MCP Integration: It acts as a Model Context Protocol (MCP) server, making the memory store accessible to clients like Claude Code, Cursor, and Cline.
• Gateway Mode: A "base station" mode allows mobile devices to access the memory store via HTTP.

Who it’s for

• Users of AI coding assistants (Cursor, Claude Code, etc.) who work across multiple devices.
• Developers building personal AI stacks who want a self-hosted, encrypted memory layer.
• Users seeking a privacy-focused alternative to cloud-hosted AI memory that avoids LLM-driven memory distillation.

Highlights

• Cross-Device Continuity: Seamlessly syncs memory across phone, PC, and tablet via encrypted cloud drive diffs.
• C-Level Integration: Native support for MCP, allowing a single memory store to be shared across various AI IDEs and agents.
• Handover Cards: A specialized memory type (kind=handover) that acts as a persistent "workbench" for long-term tasks, tracking goals, failures, and next steps.
• Zero-Config Sync: Automatic key derivation and cloud drive detection mean users don't need to manage passwords for standard synchronization.
• Edge Preloading: Pushes high-heat memories to new devices before they are opened to eliminate retrieval latency.
• Privacy-First: Implements PAMS (Privacy Access Management System) with migration tags to ensure local-only nodes never leave the device.