I made a local Markdown memory vault that OpenClaw can query over MCP

I'm the author of Synapse, an MIT-licensed local memory vault for AI agents. The first version won the LangGraph hackathon in London. I rebuilt it into something lighter because I wanted long-term context that stayed in files I could inspect and edit. Synapse imports ChatGPT exports, Claude exports, and folders of text into Markdown. SQLite FTS5 is only a rebuildable search index. OpenClaw can connect through the local stdio MCP server to search, read pages, follow graph neighbors, and propose notes for human approval. The whole vault is never automatically dumped into the prompt. Import and keyword search run locally with no model calls. Synapse has no telemetry or hosted service collecting your data. Optional wiki building sends selected source text and relevant existing pages to the OpenAI-compatible endpoint you configure. Optional semantic search sends the text you choose to embed and your search queries to your configured embedding endpoint. Quick demo with synthetic data, no API key required: uvx --from synapse-vault synapse serve --demo Repo and OpenClaw setup: https://github.com/anshulyadav1976/synapse I tested import on my own ChatGPT export: 3,138 conversations scanned, 2,044 imported in 9.3 seconds, with zero API calls. I'd like to hear how this compares with your OpenClaw memory setup, especially what you'd want an agent to remember between sessions.