Articles

ModelBrain's MCP server tools: the full list of five calls

By Matthew Firth, founder of ModelBrain · Sep 14, 2026 · 2 min read

Update, October 7, 2026: this article describes the tool surface as it stood on September 14, when it was five calls. The server now exposes fourteen tools: remember, recall, forget, expand, list_memories, six partition tools, and three read-only checks (diagnostics, usage, ping). connector_sync is no longer one of them; it is a command-line tool. The current list is on the docs page.

Here is our own proposal: keep a memory tool's call surface as small as its set of real actions. On 14 September ModelBrain's MCP server had five calls, and all three removal shapes live inside one forget call that takes a mode. The list has since grown to fourteen tools, and the additions are new kinds of action, partitions and read-only checks, not extra verbs for old ones.

Five interlocking amber module shapes in a row, with a faint dotted outline where a sixth would go

The five

remember, recall, forget, expand, and connector_sync are the entire MCP surface. Anything an assistant needs to do with your memory reduces to one of these five: save something, retrieve something, remove something, get more detail on something already retrieved, or bring a connected folder up to date.

Why not a sixth, for the thing that's obviously missing

The most tempting sixth call is something like dataset_query, for structured queries over tabular data (CSV, XLSX, JSON). It's sketched, not built, and deliberately deferred rather than bolted on early. Structured-data tooling is a genuinely different problem from document-and-assertion memory, and shipping a half-designed version of it now would mean living with an API shape decided under time pressure rather than after the ingestion formats it depends on actually exist.

The correction operations: two inside forget, one inside remember

forget takes a mode: revert (the default) brings back the version before the current one, object removes just the one version named, and chain clears a memory's whole correction history at once. All three removal shapes live in that one call (described in more detail here), because the assistant calling these tools shouldn't need a sixth or seventh tool, or to memorize which of several near-synonymous verbs does what.

A small surface is a design decision, not a limitation

Every additional call is one more thing a host has to learn correctly and one more way an assistant can misuse the API on your behalf. Five calls that map cleanly onto five real actions has held up so far. It grows only when a genuinely new action shows up that the existing five can't express, not because a bigger API looks more capable on a docs page.

Full tool reference: the docs page. How this compares to most other MCP memory servers: most store text; few tag where it came from.