Shared memory for Claude Code, Codex and Cursor
Switch from Claude Code to Codex or Cursor and the new agent can't see what the last one picked up in its sessions. Hippo keeps one store in your repo that all three use. Claude Code and Codex load from it through hooks, and Cursor reads and writes it through MCP tools. It runs on your machine, with no account and no API key.
Set up the project
$ npm install -g hippo-memory
$ hippo init
Run init in the project. It adds a short block to CLAUDE.md and AGENTS.md if they're
already there, and never creates either file. Codex and Cursor both read AGENTS.md. If the project
has AGENTS.md and Codex is installed, init adds Codex's two hooks too. For Cursor, add hippo to
.cursor/mcp.json.
{
"mcpServers": {
"hippo-memory": {
"command": "hippo",
"args": ["mcp"]
}
}
} What each agent gets
| Agent | What hippo adds |
|---|---|
| Claude Code | A block in CLAUDE.md and 9 hooks. Your pinned rules load at session start, and each prompt gets up to 5 memories that share its words. The hooks also save your task before compaction, store failed tool calls and capture lessons when the session ends. |
| Codex | A block in AGENTS.md and two hooks, which Codex runs once you trust them in /hooks. The prompt hook sends your pinned rules and up to 5 matching memories with each prompt. Capturing Codex sessions is opt-in, with hippo hook install codex. |
| Cursor | The block in AGENTS.md, which Cursor reads from the project root, and 13 memory tools once hippo is in .cursor/mcp.json. Cursor's agent recalls and stores memories through those tools, and MCP access alone doesn't capture a session. |
Where the memory lives
Each project gets its own store in .hippo/, and a global store in ~/.hippo/ holds the
memories you share across projects. Hippo writes markdown copies beside them, so you can grep what your agents
have been told, commit it and review a rule in a pull request. Recall makes no network call by default.
When a memory is wrong
After a bad answer, hippo outcome --bad weakens the memories from the last recall. hippo supersede
replaces an old fact with a new one, and the old one leaves recall but stays on record. It doesn't matter which
agent saved the memory, because they all share one store.
Import your notes and Cursor rules
$ hippo import --agents
$ hippo import --cursor .cursor/rules The first brings in the notes your agents already keep. The second turns Cursor's rules into memories that Claude Code and Codex can recall too.
What we've measured
Our repeat-mistake test ran on Claude Code only, so there's no measured result yet for Codex or Cursor. The benchmarks page has the results, losses included.
Agent questions
Which agents does hippo work with?
hippo init detects Claude Code, Codex, Cursor, OpenClaw, OpenCode, Pi and the Copilot CLI, and wires itself into each one's instruction file, hooks, plugin or MCP config. What runs automatically differs by agent. Any MCP client can use the MCP server, and anything else can call the CLI.