Self-Hosted AI Memory System v2.0

FREE drop-in replacement for Supermemory, Mem0, Zep

User message -> search relevant memories -> inject into system prompt -> AI responds -> auto-extract facts -> save

What paid tools charge $20-50/mo for:

• Vector search (semantic similarity)
• Profile extraction (static facts + dynamic context)
• Document storage & retrieval
• Deduplication

What this gives you FREE + more:

• FTS5 full-text search with BM25 ranking
• Vector search with 5 embedding options (local, OpenAI, Ollama, BM25, hybrid)
• Hybrid search (FTS + vector + importance + recency + access frequency)
• 4 memory types: static, dynamic, episodic, semantic
• 4 importance levels: critical, high, normal, low
• Memory deduplication & auto-merge
• Time-decay scoring (recent memories rank higher)
• Memory relationships (graph — link related memories)
• Auto semantic chunking for long texts
• LLM-based fact extraction from conversations
• Soft delete + restore
• Export/import, bulk ops, cleanup/maintenance
• MCP server for Claude CLI (8 tools)
• Vercel AI SDK middleware
• OpenAI middleware
• Framework-agnostic middleware (any LLM)
• SQLite storage (no cloud, no API keys)

Project Structure

D:\Betting\
|
|-- memory\
|   +-- memory-mcp-server.js      <- MCP server for Claude CLI (8 tools)
|
|-- memory-engine.js               <- Core engine (FTS5, vector, graph, dedup)
|-- embeddings.js                  <- 5 embedding providers + fact extractor
|-- memory-middleware.js            <- Auto-inject middleware (Vercel, OpenAI, generic)
|-- ai-sdk-tools.js                <- Vercel AI SDK tools (6 tools)
|-- examples.js                    <- Usage examples for every feature
|-- test-memory.js                 <- Test suite (36 tests)
|-- package.json
|
|-- .mcp.json                      <- Project-level MCP config (auto-starts server)
|-- CLAUDE.md                      <- Instructions for Claude to use memory
|-- memories.db                    <- SQLite database (created on first use)

Setup for Claude CLI

Prerequisites

• Node.js 18+
• better-sqlite3 installed (npm install better-sqlite3)
• Windows: needs Windows SDK 10.0.26100.0 for native build (VS Installer > Individual Components)

Option 1: Project-level (this project only)

Already configured. The .mcp.json file auto-registers the server:

{
  "mcpServers": {
    "memory": {
      "command": "node",
      "args": ["D:/Betting/memory/memory-mcp-server.js"],
      "env": {
        "MEMORY_DB_PATH": "D:/Betting/memories.db",
        "MEMORY_CONTAINER": "claude_memory"
      }
    }
  }
}

Option 2: Global (all projects)

Run this in a regular terminal (NOT inside Claude Code):

claude mcp add memory --scope user -- node D:/Betting/memory/memory-mcp-server.js

Verify:

claude mcp list

Option 3: JSON config (if command doesn't work)

claude mcp add-json memory "{\"command\":\"node\",\"args\":[\"D:/Betting/memory/memory-mcp-server.js\"]}"

Auto-approve MCP servers

In ~/.claude/settings.json (already configured):

{
  "enableAllProjectMcpServers": true,
  "enabledMcpjsonServers": ["memory"]
}

Claude CLI Configuration Files

8 Tools Claude Gets

Test It

Restart Claude Code in D:\Betting, then:

> My name is Ravi and I prefer dark mode
> What do you know about me?

Close and reopen:

> What's my name?

It remembers.

Run the test suite:

node test-memory.js

Expected: 36/36 PASS.

For Your AI App (not just Claude CLI)

Vercel AI SDK

import { selfHostedMemoryTools } from './ai-sdk-tools.js';

const tools = selfHostedMemoryTools({
  dbPath: './memories.db',
  containerTags: ['user_123'],
});

// 6 tools: searchMemories, addMemory, updateMemory,
//          forgetMemory, getUserProfile, getRelatedMemories

Auto-inject middleware (any model)

import { withMemory } from './memory-middleware.js';

const model = withMemory(openai('gpt-4o'), 'user_123', {
  dbPath: './memories.db',
  addMemory: 'always',
  tokenBudget: 2000,
});

OpenAI direct

import { withOpenAIMemory } from './memory-middleware.js';

withOpenAIMemory(openaiClient, 'user_123', {
  dbPath: './memories.db',
  addMemory: 'always',
});
// openai.chat.completions.create() now auto-injects memories

Any LLM (Anthropic, Gemini, Ollama, etc.)

import { withGenericMemory } from './memory-middleware.js';

const chat = withGenericMemory(myLLMFunction, 'user_123', {
  dbPath: './memories.db',
  addMemory: 'always',
});

See examples.js for 7 complete examples.

Custom Environment Variables

# Custom database path
set MEMORY_DB_PATH=D:\my-path\memories.db

# Separate memories by project
set MEMORY_CONTAINER=my_project

# With MCP registration
claude mcp add memory -e MEMORY_DB_PATH=D:/path/memories.db -e MEMORY_CONTAINER=my_project --scope user -- node D:/Betting/memory/memory-mcp-server.js

vs Paid Tools