Agent connection issues

An agent that runs without boot memory, recall, or Lore tools, or that shows Lore twice, usually has a local integration problem. Check the server and the token first: plugins fail open, so a broken connection looks exactly like Lore being turned off.

Check each runtimeCopy Anchor Link

Run the check for your runtime, and restart the runtime after any change.

No boot memory at session startCopy Anchor Link

Cause: The session-start hook did not run, or it returned nothing. The usual causes are a server or token problem, a disabled plugin, a runtime that was not restarted, or an empty runtime boot node.

Fix:

1. Ask the agent to call lore_boot. If it returns the boot view, the server and token work, and the hook is the problem. If the call fails, fix the connection first.
2. Check your runtime in the table above, then restart it.
3. For ZCode, add content to the node core://agent/zcode if that node is empty.
4. For DeepSeek Harness, add content to the node core://agent/dsh if that node is empty.
5. For OpenCode, boot needs the experimental experimental.chat.system.transform hook. If your OpenCode build does not provide it, the plugin logs one warning, skips boot, and keeps native tools and prompt recall working. The warning starts with Lore OpenCode system context is unavailable. Update OpenCode and restart it.

No <recall> block appearsCopy Anchor Link

Cause: Lore writes a <recall> block only when at least one memory passes the display threshold. The block also does not appear when the prompt hook did not run, or when recall failed, for example because no embedding model is configured.

Fix:

1. Ask the agent to call lore_search with a keyword from your prompt. If search finds the memory, the hook path works, and the problem is in recall. See Recall issues.
2. If search finds nothing either, check the server health and token, then check your runtime in the table above.
3. Test with a real question. Recall needs a prompt to search.

Lore tools are missingCopy Anchor Link

Cause: The runtime is not connected to Lore's MCP server, or the plugin did not register its tools.

Fix:

• Claude Code and ZCode: Memory tools appear as mcp__plugin_lore_lore__<tool>. In Claude Code, check with /mcp that lore is connected.
• DeepSeek Harness: Memory tools appear as mcp__lore__<tool>, for example mcp__lore__lore_search.
• Codex: codex mcp list must list lore.
• OpenCode: The native tools must use the exact names from lore_guidance through lore_move_node. A name such as lore_lore_search means an MCP entry is registered instead of the native plugin. See the next section.
• Skills tools (lore_skill_*) appear only when the server advertises Skills. The open-source server has none. Loremem Cloud does.

Lore appears twiceCopy Anchor Link

Cause: More than one integration loads Lore. The usual sources are an old user-level MCP entry, duplicate hooks, or a compatibility layer that imports Claude Code plugins.

Fix:

• Claude Code: An earlier install may have left user-scope lore or lore-skills MCP entries. Remove them, then run the install again. Skip any command that reports the server was not found. The installer scrubs the lore-skills entry itself; skills reach Claude Code through the plugin's skills/ directory now.

  Copy Text

  claude mcp remove --scope user lore
  claude mcp remove --scope user lore-skills

• Codex: Run npx @loremem/cli install --channels codex. The installer removes obsolete Lore entries from ~/.codex/hooks.json and keeps unrelated hooks. Set LORE_CODEX_INSTALL_USER_HOOKS=1 only for an older Codex build that needs user-level hooks.

• OpenCode: If you see both lore and lore:lore, or the console records two client types for one prompt, another plugin is importing Claude Code Lore hooks. Run npx @loremem/cli install --channels opencode --force. When the installer can parse oh-my-openagent.json[c] (or the legacy oh-my-opencode.json[c]) safely, it sets claude_code.plugins_override["lore@lore"] to false there. If it skips the file, set that value yourself, keep your other settings, and restart OpenCode. Do not set LORE_OPENCODE_ALLOW_MCP=1 unless you want the legacy MCP fallback alongside the native plugin.

• OpenCode, unmarked file: If the installer reports that lore-memory.js is not managed by Lore and keeps it, move that file aside and run the install again.

Behavior changed after an updateCopy Anchor Link

Cause: The runtime still runs old hooks, tools, or plugin files. Hooks and MCP config load when the process starts.

Fix:

1. Restart the runtime. In Pi, run /reload or restart.

2. Update the integrations:

   Copy Text

   npx @loremem/cli update

   update changes only integrations whose state is installed or partial. It fails instead of reporting success when an integration cannot be updated. Use --channels to pick specific runtimes.

3. To reinstall one runtime even when its version has not changed, run:

   Copy Text

   npx @loremem/cli install --channels claudecode --force

Claude Code keeps its own auto-memory. To make Lore the only memory, set CLAUDE_CODE_DISABLE_AUTO_MEMORY=1 or "autoMemoryEnabled": false in ~/.claude/settings.json. Lore never writes to CLAUDE.md. If ~/.claude/CLAUDE.md still imports @~/.claude/lore-guidance.md, delete that line.

Every Lore call failsCopy Anchor Link

Cause: The server is down, the base URL or token is wrong, or ~/.lore/config.json is malformed. The plugin hides the error from the agent, so every call looks empty.

Fix:

1. Check the server and the token with the checks on the troubleshooting page.

2. Run npx @loremem/cli status and confirm that base_url is the server you expect.

3. Run the install again with the correct server and token:

   Copy Text

   npx @loremem/cli install --base-url http://127.0.0.1:18901 --api-token lm_your_token

   Claude Code reads LORE_BASE_URL and LORE_API_TOKEN from the env block of ~/.claude/settings.json, and the installer updates them.