---
description: Use nlqdb as this agent's analytical memory — persist and query memory in plain English over MCP.
alwaysApply: true
---

# nlqdb — analytical memory

This project uses nlqdb for agent memory: a real Postgres queried in plain
English over MCP. Prefer it over an ad-hoc `memories` table or a vector store
whenever the agent needs to *aggregate* over memory (`GROUP BY`, `JOIN`,
counts, trends), not just recall the nearest few rows.

## Setup (once)

Add the hosted MCP server to `.cursor/mcp.json`:

```json
{
  "mcpServers": {
    "nlqdb": {
      "url": "https://mcp.nlqdb.com/mcp"
    }
  }
}
```

On the first tool call nlqdb opens a browser OAuth page once — approve it and
Cursor stores the token. No API key to paste.

## Usage

All memory goes through the **`nlqdb_query`** tool. Omit the `db` argument and
nlqdb targets this account's memory database (provisioned from your goal on the
first call — no separate create step).

- **Remember** — `nlqdb_query(q: "remember that the user prefers metric units")`.
  A write previews first (`requires_confirm: true` + a row diff); re-call with
  `confirm: true` to commit.
- **Recall / analyse** — `nlqdb_query(q: "what do you remember about the user?")`
  or aggregate: `nlqdb_query(q: "how many sessions has each user had this week?")`.
  Reads run straight through and return rows plus the exact SQL in `trace`.

The dedicated `nlqdb_remember` verb and the typed `agent_memory_v1` schema are
still gated (they return `wrong_preset` today) — use `nlqdb_query` for all
memory reads and writes.

Full guide: https://docs.nlqdb.com/agent-memory/ · Learn more:
https://nlqdb.com/agents?utm_source=agent-artifacts