MCP tools

SwarmMemory exposes its knowledge graph through an MCP server. This is how the Queen and your Workers read and write memories: rather than parsing files directly, they call typed tools over the Model Context Protocol. The server advertises exactly 14 tools, each with a validated input schema, giving every agent a stable, auditable contract against the graph.

Hub lifecycle

These tools initialize a graph in a repo and report on its health.

• init_hub — create the .swarmade/ memory directory and seed an empty MEMORY.md.
• hub_status — report counts (files, tags, links, broken links) and configuration such as the active embeddings provider.

CRUD on memories

The everyday read and write surface. These operate on individual Markdown files and run locally and free.

• create_memory — write a new memory file with frontmatter and body.
• read_memory — return a memory’s body, frontmatter, outgoing links, and tags.
• update_memory — patch body or frontmatter and bump the modified timestamp.
• append_to_memory — append content to an existing memory, preserving frontmatter.
• delete_memory — remove a memory file from the graph.

Discovery

Find the right memory by listing, keyword search, or meaning.

• list_memories — enumerate memories, filterable by type or tag.
• search_memories — full-text keyword search across all memories.
• semantic_search — meaning-based search using embeddings.
• list_orphans — surface files with no incoming and no outgoing links.

Linking

Keep the graph connected and navigable.

• auto_link — propose wikilinks for a memory using an LLM; returns suggestions only, never mutating the file.
• find_backlinks — list every memory that links into a target.
• suggest_connections — score related memories by shared tags and co-occurrence.

For what runs on-device versus what touches the cloud, see Local-only vs cloud sync.