First-run setup

Finish the four steps at /setup before you connect agents. You give Lore an embedding model and a chat model, then write the boot memories that every session loads.

Before you startCopy Anchor Link

• An embedding endpoint: any OpenAI-compatible /v1/embeddings service, such as OpenAI or a self-hosted model server. You need its base URL, an API key, and a model name.
• A chat endpoint for the View LLM, in one of three protocols: OpenAI-compatible Chat, OpenAI Responses, or Anthropic Messages.
• If one provider gives you both endpoints with one key, enter that key in both steps.

1. EmbeddingCopy Anchor Link

Fill in the Embedding step:

Semantic recall needs an embedding model.

2. View LLMCopy Anchor Link

The View LLM writes the gist and question views that recall searches, and it runs Dream. Pick the provider that matches your endpoint:

Enter the base URL, API key, and model for the chat endpoint. Leave the base URL empty to turn off view refinement and Dream runs.

3. Global boot memoriesCopy Anchor Link

Every session loads three global nodes. Together they are the boot baseline. Write each node as short, direct text. You can edit any node later at /memory.

core://agent: rules for every agentCopy Anchor Link

Write:

• How agents should work with you: when to read memory first, when to write a lesson back, and when to ask before acting.
• Decision defaults and boundaries that apply in every runtime.

Leave out rules for one runtime. Put those in core://agent/<runtime> in step 4.

core://soul: style and personaCopy Anchor Link

Write:

• The tone you want, such as direct, warm, or terse, and how the agent should push back.
• How the agent should behave in conversation.

Leave out task details. This node covers style only.

preferences://user: how you workCopy Anchor Link

Write:

• Stable facts that help any agent, such as your role, the languages you use, and the stack you work in most.
• Working preferences, such as "write the failing test first."

Leave out guesses, one-off task details, and secrets.

Lore's built-in usage rules are a separate setting (lifecycle.guidance.global). You don't need to copy them into a boot node.

4. Channel agent memoriesCopy Anchor Link

Add one node for each runtime you connect. Use the URI core://agent/<runtime>, for example core://agent/claudecode, core://agent/codex, core://agent/opencode, core://agent/pi, core://agent/openclaw, core://agent/hermes, core://agent/zcode, or core://agent/dsh.

Write only what applies inside that runtime:

• Hook and plugin behavior, such as how boot and recall reach the agent.
• Tool names as the runtime shows them. In Claude Code, Lore tools carry the mcp__plugin_lore_lore__ prefix, for example mcp__plugin_lore_lore__lore_search.
• Commands and conventions that apply only there.

Leave out shared rules. They belong in core://agent. Skip works here too.

5. Save and restartCopy Anchor Link

Save the settings, then restart each agent runtime you connected. Hooks and MCP clients read the server configuration when they start.

Save settings from the command lineCopy Anchor Link

Use PUT /api/settings to set the same values from a script. Send a patch object whose keys match the settings reference. Settings are stored in the database, not in .env.

curl -fsS -X PUT http://127.0.0.1:18901/api/settings \
  -H 'content-type: application/json' \
  -H 'authorization: Bearer lm_your_token' \
  --data '{
    "patch": {
      "embedding.provider": "openai_compatible",
      "embedding.base_url": "https://api.openai.com/v1",
      "embedding.api_key": "YOUR_EMBEDDING_API_KEY",
      "embedding.model": "text-embedding-3-small",
      "view_llm.provider": "openai_compatible",
      "view_llm.base_url": "https://api.openai.com/v1",
      "view_llm.api_key": "YOUR_CHAT_API_KEY",
      "view_llm.model": "YOUR_CHAT_MODEL"
    }
  }'

• Leave out the authorization header if the server has no API_TOKEN.
• Lore never displays a saved *.api_key value. It shows only whether one is set.

VerifyCopy Anchor Link

1. Check the health endpoint:

   Copy Text

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

2. Open /memory and confirm the three global nodes exist.

3. Open /recall, run a realistic query from one of your projects, and check the candidates and scores. A clear empty result is valid when no memory matches.

4. Start a new session in a connected runtime and ask the agent to call lore_boot. The boot baseline should load.

5. Send a realistic project prompt. A <recall> block should appear. Open a recalled node with lore_get_node before you rely on it.