Memo Bank Observer CLI

Observer CLI watches local coding work, asks an LLM to extract durable project knowledge, and sends the result to Memo Bank as reviewable memory suggestions.

It is designed for the simple flow:

1. Connect Observer from Memo Bank.
2. Optionally configure one LLM provider for automatic extraction.
3. Run capture or watch.
4. Approve useful suggestions in Memo Bank.

Install And Connect

To write an MCP config for a coding agent, copy the one-command setup from the Memo Bank Connect Agent wizard. The command looks like this:

npx memo-bank-observer@latest setup-mcp <agent> --key <api-key>

Supported agents are claude-code, cursor, codex, windsurf, opencode, continue, copilot, cline, and roo-code.

In Memo Bank, open Connect Agent, choose a project, and click STORE_CONFIG_LOCALLY.

Run the generated command:

npx -y memo-bank-observer@latest setup <setup-token>

This writes:

~/.memo-bank/observer.json

The file stores the Memo Bank API URL, project ID, and an API key created for Observer CLI.

Check the connection:

mb-observer status

Configure AI Extraction (Optional)

Observer can connect to Memo Bank without an LLM. Setup, status checks, dry-run context capture, and manual JSON import work immediately.

Configure an LLM when you want capture or watch to turn noisy local context into clean memory suggestions automatically. Raw diffs, summaries, and transcripts are sent from your machine to the provider you choose; Memo Bank receives only the extracted suggestion, tags, confidence score, and reason.

Run this for guided examples:

mb-observer configure-llm

OpenAI-compatible:

mb-observer configure-llm \
  --provider openai-compatible \
  --api-key <your_openai_key> \
  --model gpt-4.1-mini

Anthropic:

mb-observer configure-llm \
  --provider anthropic \
  --api-key <your_anthropic_key> \
  --model claude-sonnet-4-20250514

Gemini:

mb-observer configure-llm \
  --provider gemini \
  --api-key <your_gemini_key> \
  --model gemini-3.5-flash

Use --base-url for OpenAI-compatible providers that are not OpenAI.

Disable automatic AI extraction:

mb-observer configure-llm --provider none

Capture Suggestions

Capture the current repository diff and submit useful memory suggestions:

mb-observer capture --agent codex --summary "Implemented OAuth callback fix"

Watch a working session and capture when git state changes:

mb-observer watch --agent codex --interval 60

Include a transcript:

mb-observer capture --transcript ./session.txt --summary "Refactored billing plan logic"

Preview what the LLM would suggest without writing to Memo Bank:

mb-observer capture --dry-run --summary "Review current changes"

Capture changes against a branch or commit:

mb-observer capture --since main --agent cursor

Every submitted item goes to Memo Bank as pending_approval; users approve or reject it in the Review Queue.

Each LLM-generated suggestion includes a confidence score and a short reason, so the Review Queue can explain why the memory was detected.

Manual Import

For automation or tests, you can submit pre-extracted candidates:

mb-observer capture --from-json ./candidates.json

candidates.json:

[
  {
    "type": "decision",
    "content": "Observer CLI stores user configuration in ~/.memo-bank/observer.json.",
    "tags": ["observer-cli", "config"],
    "confidence": 94,
    "reason": "The setup path affects every local Observer installation."
  }
]

Troubleshooting

Check connection and LLM readiness:

mb-observer status

If Memo Bank is connected but LLM is not configured, run configure-llm for examples. You can also use capture --from-json ./candidates.json without an LLM.

If capture says no project ID is configured, rerun the setup command from Memo Bank.

If capture finds noisy suggestions, add a clearer --summary or use --transcript with the important session notes.