Troubleshooting

Most problems come from setup, tokens, or a runtime that wasn't restarted, not from recall thresholds. Work through the eight checks below in order, and fix the first one that fails before you change anything else.

Check the boundaries in orderCopy Anchor Link

1. Server health. Confirm that the server answers. This check needs no token.

   Copy Text

   curl -fsS http://127.0.0.1:18901/api/health

   On Loremem Cloud, check https://api.loremem.com/api/health instead. A healthy self-hosted server returns "status":"ok" and "database":"connected". If the request fails, run curl -sS on the same URL to read the response body. A 503 with "database":"disconnected" means the server cannot reach PostgreSQL.

2. Database and extensions. Confirm that the migrations ran. From the Compose directory, run docker compose logs web and look for [migrations] failed. Chinese recall also needs pg_jieba or zhparser. See Install and setup issues.

3. Setup and models. Open /setup on your server, for example http://127.0.0.1:18901/setup. The Embedding and View LLM steps must be saved, and the boot memories must exist. Recall needs a working embedding model before it can search.

4. Auth and token. If the server sets API_TOKEN, every REST and MCP request must send Authorization: Bearer <API_TOKEN>. Send this request with a valid token. A 405 response means the token was accepted. A 401 response means the token is missing or wrong.

   Copy Text

   curl -s -o /dev/null -w '%{http_code}\n' \
     -H "Authorization: Bearer lm_your_token" \
     "http://127.0.0.1:18901/api/mcp?client_type=mcp"

   Servers without API_TOKEN accept requests without a token.

5. Agent integration. Run npx @loremem/cli status and find the channel for your runtime. installed means the integration files are in place. partial or missing means the install did not finish, so run npx @loremem/cli install --channels <runtime> again. Then ask the agent to call lore_status. A working connection returns Lore online.

6. Runtime restart. Restart the agent after every install, config, hook, plugin, or token change. Hooks and MCP config load when the process starts. Pi needs /reload or a restart.

7. Memory shape. Check that the memory you expect has a focused path, one concrete disclosure, real glossary terms, and an honest priority. The Recall issues page shows how to test one prompt.

8. Recall settings. Change thresholds, weights, or timeouts only after checks 1 to 7 pass. The settings reference lists every key and its default.

Find your symptomCopy Anchor Link

What to include when you ask for helpCopy Anchor Link

Collect these details before you open an issue, and remove tokens and passwords first.

1. Versions. The installed_version line from npx @loremem/cli status. For Docker installs, also the LORE_FRONTEND_IMAGE value in ~/.lore/docker/.env.
2. Server kind. Loremem Cloud, a Docker server that the CLI manages, or another self-hosted server.
3. Status output. The full output of npx @loremem/cli status. It shows the token only as set or absent.
4. Runtime and client_type. For example claudecode, codex, opencode, pi, openclaw, hermes, zcode, dsh, or mcp.
5. The failing step. The step number from the checklist above and the exact error text.
6. For recall problems. The exact prompt, the <recall> block the agent received (or a note that none appeared), and whether the agent opened the node with lore_get_node.

If these steps do not explain the failure, search the existing issues and open a new one at https://github.com/FFatTiger/lore/issues.