Give Cursor Permanent Memory

1. Local mode does not need a key.
2. For cloud: register at aiagentsbay.com/register.
3. Open Dashboard → API Keys and copy your ab_live_... key.

Open Cursor → Settings → MCP, or edit ~/.cursor/mcp.json directly. Both formats below work.

HTTP (recommended):

{
  "mcpServers": {
    "agentbay": {
      "url": "https://www.aiagentsbay.com/api/mcp",
      "headers": { "Authorization": "Bearer ab_live_YOUR_KEY" }
    }
  }
}

Use this if HTTP is blocked. Stores your key in plaintext in the config file.

{
  "mcpServers": {
    "agentbay": {
      "command": "npx",
      "args": ["-y", "aiagentsbay-mcp@latest", "--api-key", "ab_live_YOUR_KEY"]
    }
  }
}

Tell Cursor to actually use the memory tools. In your project root, create or edit .cursor/rules/agentbay.md:

# AgentBay memory

You have AgentBay MCP connected.

- Before answering a question about this codebase, call agentbay_memory_recall
  with a query like the user's question. Use the results to inform your answer.
- After completing a real change (not a chat answer), call agentbay_memory_store
  with type=PATTERN | PITFALL | DECISION | ARCHITECTURE.
- Tag every memory with the file paths it relates to (filePaths field). This
  makes future recall code-aware.
- Don't store raw code or secrets — those belong in git, not memory.

Restart Cursor and run:

# In Cursor chat:
> What MCP tools do you have access to?

# Should mention agentbay_memory_*, agentbay_project_*, etc.
# Then test the round trip:

> Store a test memory titled "cursor verification" with content
  "AgentBay is connected to Cursor". Then recall it. Tell me which
  search strategies (alias / fts / vector / tag) found the result.

The recall result includes the per-strategy rationale recorded by phase-2 quality instrumentation. If you don't see that, check ~/.cursor/mcp.json.

File-aware recall: Cursor opens auth.ts and asks "what do we know about this file?"

# Day 1, working in Cursor on src/auth.ts
> What do we know about this file?

  agentbay_memory_recall query="auth.ts middleware"
    → No matches yet. Cursor falls back to a code read.

> The session expiry bug was in the JWT verifier — it wasn't checking
  the exp claim, only the signature. Store this as a pitfall.

  agentbay_memory_store
    title="JWT verifier missed exp claim"
    type="PITFALL"
    content="The verifier in src/auth.ts only validated signatures.
             Always check exp + nbf + iat as well."
    filePaths=["src/auth.ts"]
    tags=["auth", "jwt", "session"]
    → stored. id=k_abc

# Day 7, different chat, working on a different file
> Adding rate limits to login. Recall anything related.

  agentbay_memory_recall query="login security"
    → JWT verifier missed exp claim (PITFALL, 91% confidence)
       file: src/auth.ts
       (recalled because the file is related to "login security"
        via the auth tag and the entity graph)

> Good catch. Apply the same exp-check pattern here.

The phase-3 file claims also let multiple Cursor instances on the same project warn each other about overlapping file edits — call agentbay_task_lock with a list of files you intend to modify, and the engine reports any other agent currently editing them.