← Back to products

Vinyl

Open Source

Graph Memory (MCP)

The open-source memory MCP for AI agents

146 MCP tools across 12 capabilities
Bi-temporal knowledge graph (Kùzu)
Code intelligence (tree-sitter AST)
Ships as single Rust binary

What is Vinyl?

Reverb Vinyl is a high-performance, graph-based memory server built in Rust. It provides AI agents with persistent, structured memory through the Model Context Protocol (MCP), with 146 tools across 12 capability areas. Think of it as a brain for your agents: one that remembers everything, understands relationships, and retrieves what matters.

Architecture

Vinyl is built on three pillars: Kùzu (an embedded graph database) for relationship storage, hnsw_rs for vector similarity search, and tree-sitter for code intelligence. Everything ships as a single binary with zero external dependencies. No Docker, no databases to install. It runs alongside your IDE, compiler, and LSP on a developer laptop without resource contention.

Multi-sector memory

Vinyl organizes memory into five cognitive sectors inspired by human memory: Episodic (session-based events), Semantic (conceptual knowledge), Procedural (skills and how-to), Emotional (importance markers), and Reflective (meta-cognitive). Each sector has its own retrieval characteristics and decay patterns.

Bi-temporal knowledge graph

Every fact in Vinyl carries two timestamps: when it was recorded (transaction time) and when it was true in the real world (validity time). This enables powerful point-in-time queries like 'What did we know about the auth system on release day?'

Code intelligence

Vinyl parses your codebase using tree-sitter and builds a structural understanding, not just text search. It extracts functions, classes, interfaces, imports, and call relationships across 20+ languages, then stores them as graph entities. You can query things like 'find all callers of authenticate()', 'show classes implementing IRepository', 'list functions with no callers' (dead code), or run blast radius analysis to see what breaks if you change a function signature.

Composite scoring

Vinyl retrieves memories using a weighted formula: 0.6×similarity + 0.2×salience + 0.1×recency + 0.1×coactivation. This avoids the 'stale knowledge' problem of pure vector search and mirrors how human memory retrieval actually works.

Contradiction management

When memories conflict ('user prefers tabs' vs 'user prefers spaces'), Vinyl detects the semantic contradiction during consolidation, not just exact string matches. It supports configurable resolution strategies: newest-wins, highest-confidence-wins, or manual review, with per-entity-type merge rules.

Be the first to know when we launch

Join the waitlist for early access to Reverb Cloud and updates on the platform.