Skip to content

Tools

mnemo-mcp exposes persistent AI memory through 13 tools: 11 single-purpose memory tools (the recommended interface), the deprecated memory dispatcher that mirrors them and adds the entity-graph and point-in-time actions, and config for server status and runtime settings. No tool wraps results as external/untrusted content – memory entries are user-authored, not fetched from the open web.

Every limit parameter is clamped to 1-100.

Tool Purpose Key parameters
add_memory Store a new memory content (required), category, tags
search_memory Find memories by natural-language query query (required), category, tags, limit (default 5)
list_memories Browse memories category, limit (default 5)
update_memory Modify an existing memory memory_id (required), content, category, tags, source, importance
delete_memory Remove a memory by ID memory_id (required)
export_memories Export all memories as JSONL –
import_memories Import memories from a JSONL string or a list data (required), mode (default merge)
memory_stats Show database statistics –
restore_memory Un-archive a memory memory_id (required)
archived_memories List archived memories limit (default 5)
consolidate_memories LLM-summarize a category into a condensed memory (needs the [models.chat] cell) category (required)

Deprecated legacy dispatcher – prefer the single-purpose tools above; it will be removed in a future release. Every response carries a _deprecation notice. It covers the same operations via an action parameter, plus the actions built on the entity/knowledge-graph layer, which have no single-purpose equivalent yet.

Action Purpose Key parameters
add / list / update / delete / export / import / stats / restore / archived / consolidate Same as the single-purpose tool of the same name See the table above
search Hybrid search, with filters the single-purpose tool lacks query (required), category, tags, limit, context_type, since, until, min_importance (default 0.0), include_archived (default false)
capture Typed capture with dedup (e.g. auto-capture from a conversation) text (required, or content), context_type (default conversation), category, tags, source, importance, auto (default false)
archive_now Run the score-based archive sweep on demand –
compress Re-run LLM compression on an uncompressed memory memory_id (required)
entity_search Find memories by entity name name (required, or query), context_type = entity type filter (person, project, tool, concept, org, location, event), limit
entity_graph Knowledge-graph neighborhood subgraph around an entity entity_id or name/query (one required), depth (default 2), limit
history Timeline of memories linked to an entity entity_id (required, or memory_id)
as_of Point-in-time view of memories valid at a given timestamp as_of (required, ISO timestamp), limit (default 5)

as_of returns {memories, count, as_of}. The as_of parameter is only valid with action='as_of' – passing it alongside any other action returns an error instead of silently ignoring it, since point-in-time filtering isn’t implemented for search/list.

entity_search takes the entity type through context_type; there is no separate entity_type parameter.

Server status and runtime settings. Auth, bind address, and provider cells live in the host-owned ~/.mnemo/config.toml and are not editable through this tool.

Action Purpose Key parameters
status Show config, provider cells, auth mode, and database stats –
set Update a runtime setting (only log_level) key (required), value (required)
warmup Pre-download the local embedding model, or probe the [models.embed] cell when one is configured –
backfill_embeddings Embed memories that are missing a vector batch_size (optional, max 100)
URI / Prompt Parameters Description
mnemo://stats – Database statistics and server status (resource)
save_summary summary Prompt to save a conversation summary as memory
recall_context topic Prompt to recall relevant memories about a topic