Knowledge Base

The knowledge base stores ingested documents — markdown files, text files, web pages — chunked and indexed for retrieval. Unlike facts (which are structured, extracted knowledge), knowledge is raw reference material.

Because Moneypenny exposes its full API via MCP (Model Context Protocol), you can manage your knowledge base through natural language in any MCP-compatible client (Claude Desktop, Cursor, etc.) — or through the CLI.

Ingesting Documents

From a Local File

CLI

mp ingest path/to/document.md
mp ingest path/to/directory/

Natural Language

Ask your MCP-connected agent:

Ingest the file at docs/architecture.md into the knowledge base

From a URL

CLI

mp ingest --url "https://example.com/docs/architecture"

Natural Language

Ask your MCP-connected agent:

Ingest the page at https://example.com/docs into my knowledge base

What Happens During Ingestion

1. The document is read (or fetched)
2. Content is split into chunks (section-aware for markdown)
3. Each chunk is stored in the chunks table with FTS5 indexing
4. A documents entry tracks metadata: title, source path, chunk count

If embeddings are configured, chunks are also vectorized for KNN search.

Listing Documents

CLI

mp knowledge list

Natural Language

Ask your MCP-connected agent:

What documents are in my knowledge base?

Shows each document with its title, source, and chunk count.

Searching Knowledge

CLI

mp knowledge search "deployment pipeline"
mp knowledge search "security policy"

Natural Language

Ask your MCP-connected agent:

Search my knowledge base for security policy

Knowledge search uses FTS5 full-text matching. When combined with the unified search system (mp facts search), results from knowledge, facts, messages, and tool calls are fused together.

Knowledge vs. Facts

	Knowledge	Facts
Source	Ingested documents	Extracted from conversations or added manually
Format	Raw text chunks	Structured (content, summary, pointer)
Compression	None	Three levels (full, summary, pointer)
Versioning	None	Versioned with audit trail
Scope	Agent-local	Private, shared, or protected
Sync	Does not sync	Syncs via CRDT

Both are searchable through the same hybrid retrieval system and both contribute to context assembly during agent turns.