florinhrib

mesh-memory

An MCP server that gives agents across my machines one keyword index over shared notes.

What it does

My agents keep their notes, lessons and project records as Markdown files. mesh-memory indexes those files and lets any connected agent search and read them through three MCP tools: memory_search, memory_read and memory_sources.

How it fits together

  1. source Markdown notes The files stay the source of truth.
  2. clean gitleaks scan Before anything is stored, gitleaks scans the tree and every secret it finds is replaced by [REDACTED].
  3. index One SQLite FTS5 index Full-text search with BM25 ranking.
  4. serve MCP over streamable HTTP Every route except /health needs a bearer token. memory_searchmemory_readmemory_sources
How a note becomes searchable.

Decisions

  • Markdown stays the truth. The index sits beside the notes; it does not replace them.
  • Keyword search first. One SQLite FTS5 index with BM25 ranking, served over MCP.
  • Redact before storing. Secrets that gitleaks finds are replaced before anything reaches the index.
  • A token on every route. Only the health check is open.

Results

  • A 10-question recall test set: 9 of 10 answers in the top 5 results at build.
  • An independent security review on the day it was built; its fixes are in the code.

Limits

  • It is keyword search, not semantic search.
  • Not every agent is connected. Some are left out on purpose, one gets only the command-line client, and each machine is set up by hand.
  • It was written with Claude Code in a single session. The recall test and the security review are how I checked it.