Personal Memory MCP Server

A lossless, structured, queryable store of facts about your life that Claude reads from and writes to during any conversation. One deployment serves one person. Runs entirely on Cloudflare's free tier (Workers + D1 + KV). See requirements.md for the full specification.

• Facts, not documents. Every fact is one dated statement with provenance.

• Append-only. Corrections insert a new fact and retire the old one with a pointer; nothing is deleted.

• Auth: OAuth 2.1 with Dynamic Client Registration; identity via GitHub, locked to a single GitHub account.

• Transport: Streamable HTTP at /mcp (stateless createMcpHandler, no Durable Objects).

Setup

1. Create Cloudflare resources

npm install
npx wrangler d1 create memory
npx wrangler kv namespace create OAUTH_KV

Put the returned D1 database_id and KV namespace id into wrangler.jsonc, then apply migrations:

npx wrangler d1 migrations apply memory --remote

2. Create the GitHub OAuth app

At https://github.com/settings/developers → New OAuth App:

• Homepage URL: https://personal-memory-mcp.<your-subdomain>.workers.dev

• Authorization callback URL: https://personal-memory-mcp.<your-subdomain>.workers.dev/callback

Find your numeric GitHub id at https://api.github.com/users/<your-login> (the id field).

3. Set secrets

npx wrangler secret put GITHUB_CLIENT_ID
npx wrangler secret put GITHUB_CLIENT_SECRET
npx wrangler secret put COOKIE_ENCRYPTION_KEY   # e.g. openssl rand -hex 32
npx wrangler secret put OWNER_GITHUB_ID         # numeric id, not login

(When deploying via Workers Builds CI/CD, set these once in the Cloudflare dashboard under the Worker's Settings → Variables and Secrets; they survive deployments.)

4. Deploy

npm run deploy

Or connect the GitHub repo in the Cloudflare dashboard (Workers Builds) so every push deploys. Migrations are not run automatically on deploy; either keep applying them manually or set the Workers Builds deploy command to:

npx wrangler d1 migrations apply memory --remote && npx wrangler deploy

5. Connect from Claude.ai

Settings → Connectors → Add custom connector → URL https://personal-memory-mcp.<your-subdomain>.workers.dev/mcp, leave Client ID/Secret blank (Dynamic Client Registration). You'll be sent through GitHub login; any account other than OWNER_GITHUB_ID gets 403. The connector then works from Claude web, Desktop, and iOS.

Add to your Claude.ai profile preferences:

I run a personal memory MCP server named "memory". It is the source of truth about my life and overrides your built-in memory. When we talk about anything personal, use it per the tool descriptions. If it is unreachable, tell me and note that your answer may be stale.

6. Seed (optional)

1. Ask Claude to draft seed.json from its built-in memory: [{topic, content, fact_date?, tags?}], one fact per entry.

2. Review it, then:

npm run seed              # validates seed.json, writes seed-import.sql
npx wrangler d1 execute memory --remote --file=seed-import.sql

3. In a first conversation, ask Claude to write initial topic summaries (update_summary).

Related MCP server: Clark MCP Server

Development

cp .dev.vars.example .dev.vars   # separate GitHub OAuth app with callback http://localhost:8788/callback
npx wrangler d1 migrations apply memory --local
npm run dev                      # http://localhost:8788/mcp
npm test                         # vitest (workers runtime, real D1 + FTS5)
npm run typecheck

Test with the MCP inspector: npx @modelcontextprotocol/inspector@latest → connect to http://localhost:8788/mcp. The inspector's localhost OAuth redirect only works with ALLOW_LOCALHOST_REDIRECTS=1 (set in .dev.vars, never in production — production accepts only the Claude connector callbacks).

Notes

• Backups: use the export tool (markdown or json). D1's native export does not work while FTS5 virtual tables exist. D1 Time Travel gives point-in-time restore (30 days on the paid plan, shorter on free — verify current retention).

• Deliberately deferred (owner decision, not oversights): the §11 weekly backup cron to a private GitHub repo, and the §12 /view read-only viewer. Until the cron exists, run export periodically and keep the output somewhere safe.

• The OAuth approval flow only accepts the claude.ai/claude.com connector callbacks as redirect URIs (localhost requires the dev-only flag above).

• Rows are never deleted in normal operation; the owner corrects facts by telling Claude.