mentedb

The mind database for AI agents. TypeScript/Node.js SDK powered by napi-rs.

MenteDB is a purpose built database engine for AI agent memory. It provides vector similarity search, a typed knowledge graph, token budget aware context assembly, and cognitive features such as contradiction detection and trajectory tracking. This package delivers the full Rust engine as a native Node.js addon with zero runtime dependencies.

Install

npm install mentedb

The package ships a prebuilt native binary. If no prebuild is available for your platform, npm run build compiles from source (requires Rust and napi-rs).

Quick start

import { MenteDB, MemoryType, EdgeType } from 'mentedb';

const db = new MenteDB('./my-agent-memory');

// Store a memory
const id = db.store({
  content: 'The deploy key rotates every 90 days',
  memoryType: MemoryType.Semantic,
  embedding: embeddingFromYourModel,
  tags: ['infra', 'security'],
});

// Vector similarity search
const hits = db.search(queryEmbedding, 5);

// MQL recall with token budget
const ctx = db.recall('RECALL similar("deploy key rotation") LIMIT 10');

// Relate memories
db.relate(id, otherId, EdgeType.Supersedes);

// Forget a memory
db.forget(id);

// Close the database
db.close();

Cognitive features

CognitionStream

Monitor an LLM token stream for contradictions and reinforcements against stored facts.

import { CognitionStream } from 'mentedb';

const stream = new CognitionStream(1000);
for (const token of llmTokens) {
  stream.feedToken(token);
}
const text = stream.drainBuffer();

TrajectoryTracker

Track the reasoning arc of a conversation and predict upcoming topics.

import { TrajectoryTracker } from 'mentedb';

const tracker = new TrajectoryTracker();

tracker.recordTurn('JWT auth design', 'investigating', [
  'Which algorithm?',
  'Token lifetime?',
]);
tracker.recordTurn('Token lifetime', 'decided:15 minutes');

const resume = tracker.getResumeContext();
const next = tracker.predictNextTopics();

API reference

MenteDB

| Method | Description | |--------|-------------| | new MenteDB(dataDir) | Open or create a database at the given path. | | store(options) | Store a memory. Returns its UUID string. | | recall(query) | Recall memories via an MQL query. Returns RecallResult. | | search(embedding, k) | Vector similarity search. Returns SearchResult[]. | | relate(source, target, edgeType?, weight?) | Create a typed edge between two memories. | | forget(memoryId) | Remove a memory by ID. | | close() | Flush and close the database. |

CognitionStream

| Method | Description | |--------|-------------| | new CognitionStream(bufferSize?) | Create a token stream monitor. | | feedToken(token) | Push a token into the ring buffer. | | drainBuffer() | Drain and return the accumulated text. |

TrajectoryTracker

| Method | Description | |--------|-------------| | new TrajectoryTracker(maxTurns?) | Create a trajectory tracker. | | recordTurn(topic, decisionState, openQuestions?) | Record a conversation turn. | | getResumeContext() | Build a resume context string. | | predictNextTopics() | Predict the next likely topics. |

Types

enum MemoryType {
  Episodic, Semantic, Procedural, AntiPattern, Reasoning, Correction
}

enum EdgeType {
  Caused, Before, Related, Contradicts, Supports, Supersedes, Derived, PartOf
}

interface StoreOptions {
  content: string;
  memoryType?: MemoryType;
  embedding?: number[];
  agentId?: string;
  tags?: string[];
}

interface RecallResult {
  text: string;
  totalTokens: number;
  memoryCount: number;
}

interface SearchResult {
  id: string;
  score: number;
}

Building from source

cd sdks/typescript
cargo check          # verify Rust compiles
npm run build        # build the native addon
npm test             # run tests

License

Apache 2.0