ZenBrain Memory

@zensation/mcp

Status: early release. The tool surface is small on purpose and may still change before 1.0. The memory layers underneath it are the same ones the ZenBrain paper describes and the benchmarks measure.

An MCP server that gives any MCP client — Claude Desktop, Claude Code, Cursor, or your own — a memory that survives the conversation.

Five tools, one local SQLite file, no account and no network call.

Tool	What it does
zenbrain_store	Write something into long-term memory. Routing is automatic, first match wins: steps or wording such as "How to", "Steps:", "workflow:" make a procedure, an emotional weight above 0.5 an episode, a confidence above 0.9 a pinned core memory, anything else a semantic fact. Each call adds a memory; the same core memory stored again is updated.
zenbrain_recall	Search the episodic, semantic, procedural and core layers for what is relevant to a query; working memory only when you name it. Results come back sorted by a score from 0 to 1, each with its id and the layer it came from; a layer in which nothing matches returns recent episodes and facts, or its most successful procedures, at score 0, behind every hit, and minScore above 0 leaves them out. Every core block scores 0.5 or 0.8.
zenbrain_forget	Delete a memory for good, by the id and layer zenbrain_recall returned, together with what consolidation linked to it: forgetting an episode deletes the fact made from it, forgetting such a fact every episode consolidated into it, and both delete the working-memory copies of that content. alsoForgotten lists what went with it. A memory stored on its own and never consolidated keeps its long-term entry (duplicates on a recall result names such copies); forget a long-term result rather than a working-memory slot, which is only a copy. To correct a memory, forget it and store the corrected version.
zenbrain_consolidate	One maintenance pass: among the 100 most recent episodes, each with an emotional weight above 0.5 becomes a semantic fact, only once; working-memory slots decay, and those at a relevance of 0.01 or below are removed. Deletes nothing from long-term memory.
zenbrain_health	How full each layer is: slots in use, episodes, facts and how many are due for review, procedures, core blocks.

Install

Requires Node.js 22 or newer.

npm install -g @zensation/mcp

Configure your client

{
  "mcpServers": {
    "zenbrain": {
      "command": "npx",
      "args": ["-y", "@zensation/mcp"],
      "env": {
        "ZENBRAIN_DB": "~/.zenbrain/memory.db"
      }
    }
  }
}

Variable	Default	Meaning
ZENBRAIN_DB	./zenbrain.db	Path to the SQLite file. A leading ~/ means your home directory. A relative path is relative to the directory your client starts the server in — some clients, Claude Desktop among them, may start it in /, where nothing can be written — so give an absolute or ~/ path. :memory: gives a store that is discarded when the process exits.
ZENBRAIN_CONTEXTS	personal,work,learning,creative	Comma-separated context domains, passed to the cross-context layer. None of the five tools uses that layer yet, so the value currently has no effect.

The server speaks MCP over stdio. Stdout carries protocol traffic only; diagnostics go to stderr.

The seven layers

Storing is not filing. Which layer a memory lands in decides how it is retrieved and what consolidation does with it. The five tools reach working, episodic, semantic, procedural and core memory; short-term memory shows up only as a counter in zenbrain_health, and cross-context memory is not wired to any tool.

	Layer	Holds
7	Cross-Context Memory	Links between entities of different domains (experimental; no tool uses it)
6	Core Memory	Pinned facts
5	Procedural Memory	"How to do X" — skills and workflows
4	Long-Term Semantic	Facts, with FSRS scheduling
3	Episodic Memory	Concrete experiences and events
2	Short-Term / Session	Current conversation context
1	Working Memory	Active task focus, 7±2 items

Nothing in long-term memory decays or is deleted on its own; only working memory decays. Every fact gets an FSRS review schedule, but no tool reports a review, so through this server the schedule is not advanced. Emotional weight decides routing (above 0.5: an episode) and which episodes consolidation turns into facts. The algorithms are in @zensation/algorithms and can be read line by line.

This server configures no LLM provider, so nothing in it calls a model: routing on store is a content heuristic, and consolidation runs without generated summaries.

What this release does not do

Worth knowing before you wire it in:

• No embedding provider is configured by default, so recall is lexical, not semantic. Without vectors, recall ranks by folded token overlap — umlauts and accents are folded, rarer words weigh more, and an adjacent phrase beats the same words scattered. It will find "Kieler Foerde" in a note about the "Kieler Förde"; it will not connect "car" to "automobile". Episodes and facts are ranked over the most recent 500 of each, procedures over the 500 with the highest success rate. A layer in which the query matches nothing returns its most recent episodes and facts, or its most successful procedures, at score: 0, behind every hit; every core block comes with 0.5 or 0.8, so a block can rank above a weak hit. Working memory, when you ask for it, is ranked the same way. Pass an EmbeddingProvider through the library for true semantic matching — that is the configuration the published benchmarks were measured with.
• SQLite similarity search is a full scan. Fine for one person's memory; use @zensation/adapter-postgres for larger volumes.
• Consolidation is a tool call, not a schedule. Nothing runs it for you.
• The store is a plain file. It is not encrypted. Put it somewhere you would put a notebook. While the server runs, SQLite keeps -wal and -shm files next to it.

Using it as a library

The server factory is exported, so you can mount ZenBrain's tools on a server of your own or drive them in tests. The package is ESM; require() loads it on Node 22.12 and later (on 22.0–22.11, use import).

import { createZenBrainServer } from '@zensation/mcp/server';
import { MemoryCoordinator } from '@zensation/core';
import { SqliteAdapter } from '@zensation/adapter-sqlite';

const coordinator = new MemoryCoordinator({
  storage: new SqliteAdapter({ filename: './memory.db' }),
});

const server = createZenBrainServer(coordinator);
// connect it to any transport you like — the caller owns the coordinator's lifecycle

Zero-dependency, and where that stops

@zensation/algorithms and @zensation/core pull nothing but each other. That claim is checked in CI on every push against the packed tarballs, not against the source tree.

This package is deliberately outside that boundary. An MCP server needs the protocol SDK, so it carries one. Keeping it in its own package is what lets the core stay clean: installing @zensation/core never pulls the MCP SDK, and installing this never weakens the claim the core makes.

About ZenBrain

ZenBrain is a seven-layer, neuroscience-derived memory architecture for LLM agents, built in TypeScript and published under Apache-2.0; @zensation/algorithms and @zensation/core have no third-party dependencies. The benchmark results, and the configuration they were measured in, are reported in the paper; the reproduction packages are on Zenodo.

Works out of the box without an embedding provider — lexical ranking, nothing extra to install. In the paper, ZenBrain's benchmark runs against other systems used nomic-embed-text as the embedding provider; the retrieval ablations used OpenAI text-embedding-3-small.

• Source and issues: github.com/zensation-ai/zenbrain
• Paper: arXiv:2604.23878 · Open-access archive: 10.5281/zenodo.19353663
• Try it in the browser: zensation.ai/en/playground
• Model card: huggingface.co/zensation-ai/zenbrain
• Packages: @zensation/algorithms · @zensation/core · @zensation/adapter-postgres · @zensation/adapter-sqlite · @zensation/mcp · @zensation/ai-sdk
• Registry entry: ai.zensation/zenbrain

License: Apache-2.0