typed-recall
v0.2.0
Published
Recall — typed-edge memory MCP server for AI agents (Claude Code / Codex / Cursor / Windsurf). Wrapper that runs the Python MCP server via uvx.
Maintainers
Readme
typed-recall
The npm wrapper for Recall — a typed-edge memory MCP server for AI agents (Claude Code / Codex / Cursor / Windsurf).
This package is a thin Node.js shim that runs the Python Recall MCP server
via uvx or pipx. The actual implementation is in Python. The npm
package exists so you can install and add Recall to your MCP client with a
single command.
Install + use
Claude Code
claude mcp add recall -- npx -y typed-recallCursor / Windsurf / VS Code MCP
In your client's MCP config:
{
"mcpServers": {
"recall": {
"command": "npx",
"args": ["-y", "typed-recall"]
}
}
}Codex CLI
In ~/.codex/config.toml:
[mcp_servers.recall]
command = "npx"
args = ["-y", "typed-recall"]Standalone
npm install -g typed-recall
typed-recallWhat it actually does
On first run the shim resolves a Python invocation in this order:
uvx --from 'typed-recall[mcp,embed-bge,llm-openai]' typed-recall— uses uv to run the Python server in an ephemeral environment. Recommended.pipx run --spec 'typed-recall[mcp,embed-bge,llm-openai]' typed-recall— pipx fallback.python3 -m recall.mcp_server— assumes you already ranpip install 'typed-recall[mcp,embed-bge,llm-openai]'.
If none of these are available, it prints clear install hints.
To customize, set RECALL_MCP_CMD to a JSON array, e.g.
RECALL_MCP_CMD='["python3","-m","recall.mcp_server"]'.
To silence the launch hint, set RECALL_MCP_QUIET=1.
What Recall does
A memory layer for AI agents that stores conversational memories as a typed-edge graph, retrieves connected reasoning paths, and returns answers structurally bounded by retrieval support:
- Audit log — every memory op append-only logged
- Surgical forget — by node-id with cascading edge deprecation
- Bounded generation — non-vacuous CRC hallucination bound per answer
- Typed edges —
supports/contradicts/superseded/pivots/temporal_next - Multi-hop retrieval — HippoRAG-style entity expansion
- Sheaf-based inconsistency detection — frustrated-cycle scores
- Forman-Ricci bottleneck protection — graph topology preserved during pruning
- Local SQLite, zero cloud required
Full docs: https://github.com/yash194/recall.
Tools exposed via MCP
add_memory · search_memory · bounded_answer · forget · audit · graph_health · consolidate · stats
Requirements
- Node.js 16+
- Python 3.10+
- One of:
uv(recommended),pipx, orpip install typed-recall[mcp]already done
License
Apache-2.0.
