@memharness/mcp

@memharness/mcp

MCP server (stdio) for memharness: a bi-temporal, provenance-carrying memory primitive for AI agents. One SQLite file, no LLM or network calls in the storage layer.

It exposes seven tools to any MCP client: remember, recall (with as_of time travel), revise, diff, why, forget, stats.

Install

Claude Code:

claude mcp add memharness -- npx -y @memharness/mcp

Claude Desktop / Cursor (JSON config):

{
  "mcpServers": {
    "memharness": { "command": "npx", "args": ["-y", "@memharness/mcp"] }
  }
}

Codex (~/.codex/config.toml):

[mcp_servers.memharness]
command = "npx"
args = ["-y", "@memharness/mcp"]

The database lives at ~/.memharness/memory.db (override with MEMHARNESS_DB).

Why bi-temporal

Every fact records when it became true in the world separately from when the agent learned it, corrections supersede instead of overwriting, and every fact carries its source. So as_of can answer "what did you believe when you made this decision?", why can answer "why do you believe that?", and forget can drop everything from a given source.

Optional: hybrid recall

Recall is FTS5 keyword search by default (no model, fully offline). For semantic recall, install the optional embedding package and set MEMHARNESS_HYBRID=1:

npx -y -p @memharness/mcp -p @memharness/embed memharness-mcp

The local embedding model (BGE-small, ~130MB) is downloaded once, then offline. Stored facts are embedded automatically; there is no separate backfill step.

Optional: local usage log

Set MEMHARNESS_DEBUG=1 to append an op-name and timestamp line (never fact content) to a usage.log beside the database. Off by default, fully local.

See the main README for the full story, comparisons, and a worked example.

License

Apache-2.0