HippoDid MCP Server — persistent cloud memory for AI agents

HippoDid MCP Server
Persistent, portable memory for AI agents. Your AI forgets everything between sessions. HippoDid remembers.

HippoDid MCP Server

MCP (Model Context Protocol) server for HippoDid — connects Claude Code, Claude Desktop, ChatGPT, Gemini CLI, Codex CLI, Cursor, and any MCP-compatible client to persistent AI character memory.

What is HippoDid?

HippoDid is a cloud memory infrastructure for AI agents. It stores structured memories (preferences, decisions, skills, relationships) in a searchable database and retrieves them across sessions, tools, and AI providers. Write a memory in Claude Code, recall it in ChatGPT. Switch from Cursor to Gemini — your agent's memory travels with it.

Key features:

Cross-platform memory — same memories accessible from any MCP client or ChatGPT via OAuth
Structured extraction — raw text is automatically parsed into categorized facts with salience scores
Temporal decay — memories fade naturally by category (skills are evergreen, events fade in 14 days)
Character scoping — each agent/project gets its own memory namespace
BYOK AI — bring your own OpenAI, Anthropic, or Gemini key for extraction

Quick Start

1. Get an API Key

2. Download the JAR

Download the latest release from GitHub Releases.

3. Configure Claude Desktop

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

config.json

{
  "mcpServers": {
    "hippodid": {
      "command": "java",
      "args": ["-jar", "/path/to/hippodid-mcp-server-1.1.0.jar"],
      "env": {
        "HIPPODID_API_KEY": "hd_key_your_key_here",
        "HIPPODID_CHARACTER_ID": "your-character-uuid"
      }
    }
  }
}

4. Start Using

Ask Claude to remember things, search memories, sync files, and more.

MCP Tools

Tool	Description
`create_character`	Create a new character to store memories for
`list_characters`	List all characters accessible to you
`add_memory`	Add a memory via the AUDN pipeline (rule-based salience + dedup)
`add_memory_direct`	Write a memory directly, bypassing AUDN pipeline (Starter+ tier)
`search_memories`	Hybrid semantic + keyword search across memories
`sync_file`	Sync a local file to the HippoDid cloud
`import_document`	Import a document and extract memories (Starter+ tier)
`list_synced_files`	List files synced to the cloud for a character
`get_sync_status`	Get sync status summary for a character
`export_character`	Export all memories for a character as Markdown
`add_watch_path`	Sync a file and register the path for background tracking
`list_watch_paths`	List all watched file paths in this session
`force_sync`	Force an immediate sync of all watched paths
`configure_ai`	Configure tenant BYOK AI providers
`test_ai_config`	Test connectivity of saved AI provider configuration

Environment Variables

Variable	Required	Default	Description
`HIPPODID_API_KEY`	Yes	—	Your HippoDid API key
`HIPPODID_CHARACTER_ID`	No	—	Default character UUID for background sync
`HIPPODID_BASE_URL`	No	`https://api.hippodid.com`	API base URL
`MCP_SYNC_INTERVAL`	No	`300`	Background sync interval in seconds
`MCP_AUTO_CAPTURE`	No	`false`	Disable background watcher (AI handles sync)
`MCP_AUTO_RECALL`	No	`false`	Disable background hydration (AI handles download)
`MCP_RECALL_CACHE_TTL`	No	`120`	Search result cache TTL in seconds

Architecture

Code

Claude Desktop → stdio (JSON-RPC) → McpServerRunner → HippoDidClient → HTTP → HippoDid REST API

The MCP server uses the hippodid-spring-boot-starter to communicate with the HippoDid REST API over HTTP. All authentication and tenant isolation is handled server-side via Bearer tokens.

Building from Source

Requires Java 21+ and Maven 3.8+.

bash

mvn clean package -DskipTests
java -jar target/hippodid-mcp-server-1.1.0.jar

Versioning

The MCP server version matches the hippodid-spring-boot-starter version. Both must be at the same version for compatibility.

License

Apache License 2.0 — see LICENSE.