Troubleshooting

Troubleshooting starts by finding the failing boundary.

Do not tune recall thresholds first. Most problems are caused by setup, auth, stale hooks, incomplete model settings, or weak memory shape.

Check in this order

1. Server health — /api/health responds and the Web Console loads.
2. Database and migrations — DATABASE_URL is correct; PostgreSQL has vector; Chinese recall has pg_jieba or zhparser when needed.
3. Setup — /setup reports complete embedding, View LLM, and boot memory setup.
4. Auth — protected servers receive Authorization: Bearer <API_TOKEN> from REST and MCP clients.
5. Agent connection — MCP URL includes the right client_type; plugin files and hooks are installed.
6. Runtime reload — the agent was restarted after config, hook, or plugin changes.
7. Memory quality — path, disclosure, glossary, and priority describe the future prompt clearly.
8. Recall settings — only tune thresholds after the above checks pass.

Use the focused pages

Keep troubleshooting evidence concrete: URL, client_type, token state, page name, query text, recalled URI, and whether the agent opened the memory node.

Uninstall

To fully remove Lore from your machine:

Shell

# Stop and remove containers, networks, and volumes
docker compose down -v

# Remove local Lore state
rm -rf ~/.lore

# Remove Claude Code Lore integration
rm -f ~/.claude/lore-guidance.md
# Remove the @~/.claude/lore-guidance.md import line from ~/.claude/CLAUDE.md

# Remove Codex MCP registration
codex mcp remove lore || true

# Remove OpenClaw plugin entry from ~/.openclaw/openclaw.json

# Remove Pi extension
rm -rf "${PI_CODING_AGENT_DIR:-$HOME/.pi/agent}/extensions/lore"

# Remove Hermes plugin symlink
rm -f ~/.hermes/skills/lore_memory

Also remove LORE_BASE_URL and LORE_API_TOKEN from your shell profile if you added them.