Skip to content
AG
← All Projects

Memory MCP Server

Memory MCP Server

Overview

A production MCP server that gives AI assistants persistent memory across conversations. Features 8 tools — save, search, update, delete, list, get_context, get_rules, and get_memory. Uses pgvector with gte-small embeddings (384-dim) for AI-powered semantic search across all stored memories. Secured with OAuth 2.0 + PKCE (S256). Includes a glassmorphism UI with a visual rules editor that exports .claude.md files. Soft deletes, access logging, and category-based organization (preference, rule, project, decision, context, snippet, note, persona).

Tech: TypeScript, Next.js 16, React 19, MCP Protocol, Supabase, pgvector, Tailwind v4, ShadCN, OAuth 2.0, Vercel

Tech Stack

TypeScriptNext.jsReactSupabaseTailwind CSSVercel

Key Features

8 MCP tools for full memory CRUD + semantic search

AI-powered vector similarity search via pgvector + gte-small embeddings

OAuth 2.0 with PKCE S256 for secure AI client auth

Category-based memory organization (preference, rule, project, decision, context, snippet, note, persona)

Visual rules editor with .claude.md export

Glassmorphism UI with indigo accents built on ShadCN v4

Soft deletes and access logging on every tool call

Project-scoped context retrieval for instant AI onboarding

Challenges & Solutions

Semantic Search Accuracy

Implemented pgvector with gte-small embeddings (384-dim) via Supabase Edge Functions, enabling natural language queries to find relevant memories with high precision.

Secure AI Client Authentication

Built full OAuth 2.0 flow with PKCE S256 challenge method, allowing AI clients like Claude to authenticate securely without exposing secrets.

Cross-Session Context Continuity

Designed project-scoped get_context tool that retrieves all related memories in one call, giving AI assistants instant onboarding into any project's full state.

Architecture

Next.js 16 App Router with Server Components by default. MCP server exposes 8 tools via MCP SDK. Supabase handles auth, data storage, and vector embeddings via Edge Functions. pgvector powers semantic search with 384-dim gte-small embeddings. OAuth 2.0 with PKCE S256 for secure AI client authentication. Deployed on Vercel.