Skip to content
Pathrule

compare

Pathrule vs the MCP memory server

The reference memory server gives an agent a local knowledge graph of entities, relations and observations, and eight tools to write and read it. Pathrule speaks the same protocol, so the difference is not transport. It is that the reference server waits to be called, keeps one shared graph with no notion of where you are working, and cannot enforce anything.

What the reference server is for

It is the canonical example of an MCP server that holds state, and it is genuinely the fastest way to understand what MCP is. Entities carry observations, relations connect them in active voice, everything persists to a JSONL file you can open in an editor, and the whole thing runs from a single command with no account and no service.

For a solo experiment, for learning the protocol, or for a handful of facts you want an agent to remember across conversations, that is enough and nothing here argues otherwise.

Side by side

Reference MCP memory serverPathrule
TransportMCPMCP, plus hooks and instruction files installed by the CLI or the VS Code extension
How knowledge arrivesThe agent must call a tool: `search_nodes`, `read_graph`Delivered before the first tool call through the Pathrule hook in Claude Code and Codex
ScopeOne shared graph, no path or user scopingIndexed by repository path, per workspace
What is storedEntities, relations, observationsMemories, rules and skills, typed and prioritised
EnforcementNoneA strict rule can block a matching change where the Pathrule hook runs (Claude Code and Codex)
TeamA local JSONL fileShared workspace with roles and review
Stale knowledgeYours to noticeRaised as a proposed edit you approve

Reference server claims are from its README in modelcontextprotocol/servers, read August 2026. If a row is wrong, tell us and we will fix it.

The differences that follow from that

  • A tool the agent must remember to call. Its README is explicit that agents must call the tools; nothing is automatic. So the agent has to decide to search before it acts, which is the same gap every query-driven memory has: the constraint you most need is the one nobody thought to look up.

  • One graph, no sense of place. There is no repository-path scoping and no isolation: it is a single shared graph for every interaction. That is fine for a few dozen facts and unworkable once two services in one monorepo disagree about what a valid change looks like.

  • Nothing can be a constraint. Observations are text an agent may read. A Pathrule rule carries a scope, a priority and an enforcement level, so in Claude Code and Codex the local Pathrule hook can stop a matching change before it lands rather than leaving it to the model's judgement.

When the reference server is the right choice

Use it when you are learning MCP, when you want a dependency-free scratchpad for one agent on one machine, or when the knowledge really is a handful of entities rather than a team's accumulated conventions. It is a reference implementation doing exactly what a reference implementation should.

Reach for Pathrule when the knowledge has to be scoped to where the work is happening, when it has to reach a teammate's session and a second engine, when some of it must gate a change rather than suggest against one, and when you would rather the agent proposed what to keep than maintain a graph by hand.

Sources and verification

Official reference-server documentation verified 1 October 2026. No released reference-server version was locally tested. Pathrule claims reviewed against the Pathrule v1 CLI on 19 October 2026.

Frequently asked questions

Pathrule is an MCP server too. What is actually different?

The protocol is the same; the delivery is not. The reference server hands over knowledge when an agent calls a tool. In Claude Code, and in Codex when the repository has Codex files, pathrule setup also registers hooks and writes instruction files, so the slice that governs the path you are working in arrives before the first tool call, whether or not the agent would have asked. Cursor, GitHub Copilot and Windsurf get the MCP entry only, so for them, as with the reference server, knowledge arrives when the agent calls a tool.

Can I use the reference memory server alongside Pathrule?

Nothing stops you, but they overlap: both are places an agent keeps knowledge. Running both usually means two half-maintained stores, and the one with no path scoping is the one that decays. Pick per project.

Does Pathrule work if I can only use a hosted MCP endpoint?

Yes, with one limitation worth knowing. The CLI, the VS Code extension and Remote MCP are ways into the same Pathrule knowledge. Remote MCP at https://cloud.pathrule.io/mcp exposes the same workspace, memories, rules and skills with OAuth and a Pathrule account, but it does not install hooks or write local files, and it has no local mode, so it answers on request rather than delivering before the first tool call. Hook-time delivery comes with the CLI or the VS Code extension.

Does Pathrule store entities and relations like a graph?

It keeps a knowledge graph, but the index is the repository path rather than an entity name, and the content is typed: memories are facts, rules are constraints with scope and priority, skills are named procedures. That typing is what makes enforcement and ranking possible.

Get started with Pathrule.