Troubleshooting
If r3 is not working as expected, start with the smallest failing step: installation, MCP connection, storage initialization, write, read, then restart. This page covers the common symptoms and fixes.
Installation fails on macOS with a GNU Make error
The embedded Redis dependency compiles from source when a prebuilt binary is not available. This needs GNU Make 4 or later, which is newer than the system default on some macOS versions.
Install a newer make with Homebrew:
brew install makeConfirm the version:
gmake --versionThen relaunch r3:
npx @n3wth/r3MCP client does not discover r3 tools
- Restart the MCP client after editing configuration.
- Confirm the config file path matches your client (Claude Desktop, Claude Code, or Cursor).
- Verify the JSON contains
"command": "npx"and"args": ["@n3wth/r3"]. - Check the client’s MCP log for stderr output. Add
"DEBUG": "true"to the server’senvfor a diagnostic run. - Confirm the configured package is
@n3wth/r3, not an unscopedr3name.
Keep stdout reserved for MCP messages. Disabling quiet mode can let local logging reach stdout and interfere with the transport.
add_memory returns Saved but search finds nothing
- Confirm the intelligence mode. Try
INTELLIGENCE_MODE=basicto rule out embedding model initialization failures. - Call
sync_statusto check pending async operations. - Call
optimize_cacheto refresh the cache and index. - Verify you are querying with the same
user_idyou used for the write.
Saved can return before async processing finishes. Verify content through retrieval, not through the acknowledgement.
Memories do not persist across restart
- Restart from the same working directory. The vector index lives at
./data/vectra-indexrelative to the launch directory. - Local Redis records have a 30-day TTL. The embedded backend does not configure durable persistence.
- Keep an independent copy of anything you cannot lose. Test the write, read, and restart cycle with disposable data first.
See Configuration for storage details.
Redis connection refused
This applies only when REDIS_URL points at an external Redis. Ensure that server is running and reachable from the r3 process.
For the local quickstart, leave REDIS_URL unset so r3 uses the embedded server.
Enhanced-mode tools are missing
extract_entities, get_knowledge_graph, and find_connections only register when INTELLIGENCE_MODE=enhanced (the default).
If they are missing, confirm you have not set INTELLIGENCE_MODE=basic, then check stderr for model download or embedding initialization errors. On failure the server continues in a degraded state and these tools stay hidden.