Komplyt Zero MCP server

Komplyt Zero

An open-core memory graph for your AI tools. Run it yourself, or use ours in the EU.

Agents lose the reasoning behind a decision when the session ends, and each client keeps separate notes. Zero is an MCP server that gives Claude, ChatGPT, Cursor and any other MCP client one shared memory. Agents record decisions, problems and approaches together with the reason, as typed nodes in a graph. A later session, in any client, finds them by meaning instead of keyword, so the reasoning behind last month's choice is there when someone is about to undo it. It is for developers who use several AI clients on the same product, and for teams that keep re-deciding things because nobody remembers why.

A notes file grows until nobody reads it, and a plain vector store returns the closest text with nothing around it. Zero keeps the reason with each decision, links related nodes so a hit brings its context along, and lets knowledge nobody uses drop down the ranking. Nothing is deleted. It stores what the agent was asked to keep, not your chats.

The core is Apache-2.0. The team features in src/ee/ are source-available: free for noncommercial use under PolyForm Noncommercial, with a commercial licence for everything else. Zero also runs on SurrealDB (Business Source License), so the repository as a whole is not OSI open source. Details are in Licence.

To try the hosted service (free for one person, data stored in Frankfurt), add it to Claude Code and run /mcp to sign in. To run it on your own machine, see Getting started.

claude mcp add --transport http komplyt-zero https://kg.kompl.yt/mcp

What it looks like

An agent working on a webhook service records a decision:

{
  "tool": "kg_node_create",
  "arguments": {
    "type": "decision",
    "project": "acme-api",
    "tags": ["architecture", "webhooks"],
    "content": "Webhook delivery uses a Postgres job table (SKIP LOCKED) instead of Redis or SQS. Why: under 1k jobs/min, Postgres already runs, and enqueueing in the same transaction as the business write means no lost events. Revisit above 10k jobs/min."
  }
}

Three weeks later, in a different client, someone asks a question that shares no keywords with it:

{
  "tool": "kg_query",
  "arguments": { "text": "why don't we use Redis for background jobs?", "project": "acme-api" }
}

The decision comes back first. Spreading activation also returns the nodes linked to it, such as the open problem about retry storms that someone related to it last week.

How it ranks what comes back

Every node carries an embedding and a decay score. A query ranks candidates by 0.7 × cosine similarity + 0.3 × decay score, then follows the graph's edges for up to three hops (two by default) to pull in related nodes, each hop weakening the signal by half.

The decay score is how the graph forgets. Each time a session starts in a project, every node in that project keeps 85 % of its score, so scores fall off exponentially. Nodes that agents update climb back toward a ceiling of 0.85, by more when they sit in a dense neighbourhood, and a query gives a smaller lift. Knowledge that keeps being used stays near the top, and knowledge nobody touches fades without anyone having to curate it. Decay never deletes anything; a faded node still turns up when a query matches it closely enough.

The graph, the decay and the spreading activation are all in this repository under Apache-2.0. The hosted service runs the same ranking code.

Getting started

Shortened. The full README is on GitHub.