AI and LLMAgents

Memory for AI Assistants over MCP

Give Claude, Cursor, VS Code or your own agent a memory, notes, an event log, prompt templates and document search, through one MCP server.

fixed tools: 28
arbitrary commands: 0
open source: Apache-2.0

Step 1 of 5

Connect the server

One command adds HotKV to Claude Code. Cursor and Claude Desktop use an mcpServers entry instead.

Example sessionplaying

# Connect the server

$ claude mcp add hotkv -- uvx hotkv-mcp --url hotkv://localhost:6379 --namespace team-docs

(hotkv added as an MCP server)

An animated walkthrough with real HotKV commands and example replies. Click any step to jump to it.

The problem

AI assistants forget everything when the chat ends, and giving them raw database access to fix that is dangerous. One wrong command from a model can flush data or read another tenant's keys.

How HotKV handles it

hotkv-mcp gives an assistant a fixed set of 28 tools for keys, hashes and lists, conversation memory, working notes, an event log, rolling summaries, a prompt registry and document search. There is no tool that runs an arbitrary command, a namespace confines the agent to one prefix, and read-only mode removes every write tool.

Step by step

What happens, command by command

01
Connect the server
One command adds HotKV to Claude Code. Cursor and Claude Desktop use an mcpServers entry instead.
claude mcp add hotkv -- uvx hotkv-mcp --url hotkv://localhost:6379 --namespace team-docs
02
"Remember our release checklist"
The assistant saves a working note with the memory_set_note tool. It lands under the team-docs prefix.
tool memory_set_note {"session": "releases", "key": "checklist", "value": "1. bump version 2. changelog ..."}
03
Next week, in a new chat
A fresh conversation has no history, but the note is still in HotKV.
tool memory_get_note {"session": "releases", "key": "checklist"}
04
Search the team's documents
Documents ingested with rag_ingest are searched by keyword relevance, no embedding model needed.
tool rag_search {"pipeline": "runbooks", "query": "rotate TLS certificates", "k": 3}
05
Other teams stay out of reach
Every name is stored under team-docs:, so this targets team-docs:other-team:secrets, which does not exist. Another team's keys cannot even be named, and no tool flushes the database or runs raw commands.
tool delete {"key": "other-team:secrets"}

Where it fits

Real-world scenarios

Engineering teams

A shared, read-only MCP server lets every developer's assistant search the same runbooks and design notes.

Personal assistants

Preferences, ongoing tasks and decisions are kept as notes and summaries, so the assistant picks up where it left off.

Agent platforms

Each customer's agent gets its own namespace on one server and cannot see the others.

Built in

What you get

A fixed, safe tool set

28 tools and no way to run an arbitrary command. Each tool declares whether it reads, writes or deletes.

Namespaces

--namespace confines an agent to one prefix that it never sees. Give each agent a prefix that is not the start of another's.

Read-only mode

--read-only does not even register the write tools.

Bounded results

Values and lists are capped so a huge value cannot flood the model's context.

Claude Desktop, Cursor and other mcpServers clients

json

{
  "mcpServers": {
    "hotkv": {
      "command": "uvx",
      "args": ["hotkv-mcp", "--read-only"],
      "env": { "HOTKV_URL": "hotkv://:your-password@localhost:6379" }
    }
  }
}

Reference:hotkv-mcp on GitHub

Common questions

Which clients work with it?

Any MCP client, including Claude Code, Claude Desktop, Cursor and VS Code. It speaks stdio for local clients and streamable HTTP for shared use.

Does it need a HotKV license?

The key, hash and list tools work on any server. The memory, prompt and document search tools use HotKV's AI engines, which need a license that includes them.

Is the HTTP transport authenticated?

No. Keep it on localhost or behind a reverse proxy that authenticates, and give it its own namespace and read-only mode where you can.