plumb

name	plumb
description	Persistent cross-session memory for OpenClaw agents — automatically ingests conversations and injects relevant memories before each response.

Plumb — Cross-Session Memory Plugin

Plumb gives OpenClaw agents long-term memory that persists across sessions. It stores conversation facts in a local SQLite database and retrieves relevant context before each response.

Setup

1. Install the Plumb packages (from the monorepo root):

pnpm install && pnpm build

2. The plugin connects to the @plumb/mcp-server binary via stdio. By default it looks for plumb-mcp on $PATH. To use an explicit path:

{
  "mcpServerPath": "./node_modules/.bin/plumb-mcp",
  "userId": "default",
  "enabled": true
}

3. The plugin hooks into OpenClaw's message pipeline:
   • After every exchange — ingests the conversation into the memory store
   • Before every response — searches memory for relevant context and injects it

How It Works

• Layer 1 (Raw Log): Every conversation chunk is stored verbatim with vector embeddings for semantic search.
• Layer 2 (Facts): An extraction step distills conversations into structured subject-predicate-object triples with confidence scores and decay rates.
• Search: Queries both layers in parallel, combining semantic similarity with recency and confidence scoring.

MCP Tools

The underlying MCP server exposes four tools:

Tool	Description
memory_store	Persist a fact or conversation chunk
memory_search	Query both layers by semantic similarity
memory_delete	Soft-delete a fact by ID
memory_status	Get store statistics (counts, last ingestion, size)