mem-mcp documentation

mem-mcp is a personal memory layer that lets multiple AI assistants (Claude Code, Claude.ai, ChatGPT, Cursor) share the same long-term memory you control. It speaks the Model Context Protocol — so you connect each assistant once, and they all see and write to the same memory store.

Quick start

1. Sign in with Google. Your access must already be approved by the operator.
2. Connect at least one MCP client (instructions below).
3. Try it: ask your assistant to “remember that I prefer dark mode”, then in another session ask “what are my preferences?”.

The MCP endpoint

All clients connect to the same endpoint:

https://memsys.dheemantech.in/mcp

Authentication is OAuth 2.0 with Dynamic Client Registration (DCR). Most MCP clients handle this automatically — you click “connect”, sign in with Google, and they store the access token.

Connecting Claude Code

Add this server to your Claude Code config. In a terminal:

claude mcp add --transport http memsys https://memsys.dheemantech.in/mcp

Then restart Claude Code. The first tool call will trigger the OAuth login flow in your browser. After authorizing, Claude Code remembers the token.

Connecting Claude.ai (chat)

Open claude.ai/settings/connectors, click Add custom connector, and use:

• Name: memsys
• URL: https://memsys.dheemantech.in/mcp

Click connect, sign in with Google, and the connector becomes available in chat.

Connecting ChatGPT (Developer Mode)

In ChatGPT, open Settings — Developer Mode — MCP Servers (requires a ChatGPT plan that includes Developer Mode). Add the same URL: https://memsys.dheemantech.in/mcp. Authorize via Google.

Available tools

Once connected, the assistant sees these tools. Most are auto-discovered and invoked when the conversation needs them.

Memory tools (core)

• memory_write — store a note, decision, fact, snippet, or question
• memory_search — hybrid retrieval (semantic + keyword, recency-decayed)
• memory_get — fetch one memory by id (+ optional version history)
• memory_list — browse memories by type / tag / date
• memory_update — edit content or tags in-place
• memory_delete — soft-delete (recoverable for 30 days)
• memory_undelete — restore within the 30-day grace window
• memory_supersede — replace a decision/fact with a new version while preserving history
• memory_thread_get — fetch a root memory plus all its replies in order
• memory_export — full data dump for backup / portability
• memory_feedback — record explicit thumbs up / down on a search result
• memory_stats — counts, quota usage, embeddings consumed today

Onboarding

• memsys_enable_kite — wire your Zerodha Kite Connect credentials to enable broker tools (skill framework, opt-in per tenant)

Kite skill (optional, broker integration)

Enabled after a successful memsys_enable_kite call. All operate against your own Zerodha account; nothing is shared across tenants.

• kite_get_holdings — long-term equity holdings
• kite_get_positions — open intraday + F&O positions
• kite_get_margins — available cash and used margin
• kite_get_quote — live LTP / depth for one or many instruments
• kite_get_historical_data — OHLC bars for charting / pattern analysis
• kite_get_orders — today's order book
• kite_place_order — submit a new order (use with care)
• kite_cancel_order — cancel a pending order

Tags, types, and search

Every memory has a type (one of: note, decision, fact, snippet, question) and an optional list of tags. Search ranks by hybrid semantic + keyword relevance and decays older items. Tags are exact-match filters; combine them with a query for sharp retrieval.

Privacy & data

See the privacy policy for what we collect, where it lives, and how to export or delete your data. Short version: AWS Mumbai, encrypted at rest, no third-party analytics, you can export or delete everything at any time.

Source & status

mem-mcp is in closed beta. Bugs and feature requests go to the operator — right now that's a person, not a form. Email anand@dheemantech.com.