A memory layer that follows you, not your apps

What it does

MemoryBridge is a Python implementation of CDSMP (cross-device semantic memory continuity): one semantic memory store that syncs across your devices via a cloud-drive channel (Jianguoyun WebDAV primary, OneDrive backup), and is shared across MCP clients like Claude Code, Cursor, and Cline. It stores memories as a graph with typed edges and evidence, retrieves via vector + keyword + graph fusion (RRF), and can inject relevant memories into your sessions automatically via hooks. An old phone can run as a 5–10W always-on gateway so mobile devices read and write the same store over HTTP.

The interesting bit

The design’s core principle is “content freeze”: the system only extracts semantic links and tunes structural parameters — it never rewrites or summarizes your original memories, citing Faulty Memory research that LLM-rewritten memory inevitably introduces hallucinated distortion. Truncation under token budget is allowed; rewriting is not. The handover-card mechanism (a kind=handover note that becomes a persistent “workbench” across devices, auto-expiring after 7 days) is a genuinely thoughtful take on task continuity.

Key highlights

• Incremental diff-based sync with per-device watermarks and idempotent, content-fingerprinted packages
• Edge preloading: hot memory clusters are pushed to a device before you open it, so switching is instant rather than retrieval-lagged
• Schema fingerprinting and peer reconciliation so heterogeneous ends can detect and repair drift
• Human-readable Markdown export of the whole store — read-only, auditable, Git-friendly
• Core is zero-dependency (pure standard library), Python 3.9+

Caveats

• The “zero-config” channel key is deterministically derived from a seed published verbatim in the README — anyone (or any AI) can compute it without installing the project. That’s a deliberate tradeoff for convenience, but it means the channel’s confidentiality rests on the algorithm being public, which is an unusual threat model to say the least.
• The feature table is long and version-dense (v0.9 through v0.33); the README reads more like a changelog than an introduction, so expect some archaeology to find the actual architecture docs (the RFCs linked at the top are the better entry point).

Verdict

Worth a look if you work across multiple machines and MCP clients and want memory continuity without a cloud vendor in the loop. If you wanted a single-box memory layer like Mem0 or Letta, this is more machinery than you need — and you should read the threat model before trusting the key derivation.