What is analytical memory for an AI agent?

A real, queryable database an agent uses to remember — typed rows it can GROUP BY, JOIN, and aggregate over in plain English, not an opaque vector blob it can only search. Recall is table stakes; analysis over memory is the part a vector store structurally can't do.

How is this different from a vector store or vector database for agent memory?

A vector store returns the top-k most similar chunks — semantic recall with no query planner, so rollups become the model doing arithmetic over search hits. nlqdb stores typed rows in Postgres, so it answers “how many times did this user ask about pricing this month?” with one GROUP BY.

Does my agent have to design a database schema first?

No. The database is created from a goal, not a schema. Your agent writes typed rows as it learns and asks questions in English; nlqdb provisions the Postgres database and infers the shape.

Is it safe to let an AI agent write SQL against its memory?

The LLM never emits SQL. It returns a typed JSON plan; nlqdb's compiler emits parameterised SQL, an AST re-parse checks it against a verb and table allow-list, and destructive writes show a diff you confirm first — and every query returns the exact SQL it ran.

Can my agent run similarity or vector recall too?

Similarity search over an embedding column (pgvector) is an opt-in slice still landing. Today the wedge is the analytical side — the GROUP BY, JOIN, and per-group aggregate queries a vector store can't run.

What does it cost, and can I use my own LLM key?

Bring your own Anthropic, OpenAI, Gemini, or OpenRouter key at 0% markup on the free tier — no card, no per-call fees. The engine, CLI, MCP server, and SDKs are source-available under FSL-1.1 and self-hostable for any non-competing use.

Memory your agents can `GROUP BY`.

Give your agent a real database as its memory — typed rows it writes as it learns, queried in English. Not the top-k a vector store returns.

Spin up a database from a sentence.

Anonymous — no sign-in. Your DB lives 72h; sign in (always free) to keep it.

Pre-loaded with an agent-memory goal — edit it, or paste your own.

Install in your agent.

Run in your terminal.

claude mcp add --transport http nlqdb https://mcp.nlqdb.com/mcp

Add to ~/.codex/config.toml.

[mcp_servers.nlqdb]
url = "https://mcp.nlqdb.com/mcp"

Settings → Connectors → Add custom connector — paste the URL.

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

Cascade → MCPs → Configure, or ~/.codeium/windsurf/mcp_config.json.

{
  "mcpServers": {
    "nlqdb": {
      "serverUrl": "https://mcp.nlqdb.com/mcp"
    }
  }
}

Agent Panel → Add Custom Server, or ~/.config/zed/settings.json.

{
  "context_servers": {
    "nlqdb": {
      "url": "https://mcp.nlqdb.com/mcp"
    }
  }
}

Claude Code: Run in your terminal.
Codex: Add to ~/.codex/config.toml.
Claude: Settings → Connectors → Add custom connector — paste the URL.
Windsurf: Cascade → MCPs → Configure, or ~/.codeium/windsurf/mcp_config.json.
Zed: Agent Panel → Add Custom Server, or ~/.config/zed/settings.json.

Anonymous — the configs ship with the pk_live_REPLACE_ME placeholder. Sign in (free) to inline your live key →

Two MCP tools. One database.

Your agent writes a row, then asks a question.

Over MCP your agent calls nlqdb_remember to write a typed row, then nlqdb_query to ask in English. nlqdb compiles the question to SQL through a typed-plan trust boundary and runs it on Postgres.

1 · Remember

nlqdb_remember({
  category: "user_preference",
  content: "prefers concise replies",
  user_id: "u_412"
})

→ { id: 4129, ok: true }

The row lands in agent_memory — typed columns, not an opaque vector blob.

2 · Recall

nlqdb_query({
  q: "feedback per user, ranked by count this month"
})

Compiled SQL (the exact statement nlqdb ran):

SELECT user_id, COUNT(*) AS feedback
  FROM agent_memory
 WHERE category = 'user_preference'
   AND created_at >= date_trunc('month', now())
 GROUP BY user_id
 ORDER BY feedback DESC;

3 · The answer — one GROUP BY

user_id	feedback
u_412	14
u_207	9
u_388	6
u_104	4
u_551	2

Rendered from a fixture through the same typed-plan compile path the product runs — server-rendered so it stays crawlable and works with no JS. Numbers are illustrative.

Recall is table stakes. Only one of these can aggregate.

Capability	Mem0	Zep	Letta	nlqdb
Remember a fact ("Alice has a $50k deal") Storing a fact is table stakes — every memory layer does this.
Recall facts by similarity / relevance Mem0/Zep/Letta rank by hybrid/semantic search; nlqdb recalls via SQL filters (embedding hybrid still E-05).
Top-N by value ("top 5 deals by size") Needs ORDER BY + LIMIT over the full set, not a top-k similarity search.
Aggregate per group ("average deal size per stage") A vector/graph store returns matches; the LLM would have to do the arithmetic. nlqdb runs GROUP BY in Postgres.
Time-window analytics ("deals closing this month") Mem0/Zep can filter or point-in-time-recall by time; neither aggregates across a window.
Full GROUP BY / JOIN / HAVING over memory The core wedge: a real query planner over the agent's own data.
Agent designs its own schema nlqdb provisions Postgres from the agent's first goal; the others impose a fixed memory shape (Graphiti custom entity types ≠ agent-designed tables).
Diff preview before destructive writes DDL/DML is previewed and confirmed before it applies (GLOBAL trust-UX).
Self-hostable Mem0/Letta are Apache-2.0 self-hostable; Zep self-hosts Graphiti but the platform is hosted; nlqdb is FSL source-available (GLOBAL-019), container pull-forward still WS-11.

Five boxes — or one MCP tool.

Vector store top-k
Per-agent SQLite
Embedding cache
Schema design
Re-rank pipeline

 nlqdb_query(…)

Connect your agent →

Connect over MCP.

nlqdb is a hosted MCP server. Add it to Claude, Cursor, Codex, or any MCP host — OAuth opens in your browser on the first tool call, no key wired up front. No browser in the loop? Run the same server locally with a key.

1 · Paste the connector URL

Into your host’s MCP config — Claude Desktop, Cursor, Codex, Zed, Windsurf.

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

Or: one command (Claude Code)

One line adds the hosted server; one-click installs for Cursor, VS Code, and more on integrations.

claude mcp add --transport http nlqdb https://mcp.nlqdb.com/mcp

Full MCP setup →

Or: no browser at run time (headless agents, CI)

The hosted route ends on an OAuth consent screen, so an agent running in CI or on a server can’t finish it. Run the same MCP server locally instead — same five tools, authenticated by a key instead of a consent click. Mint an sk_mcp_ MCP key once at API keys (one signed-in visit) — scoped to MCP, revocable on its own — paste it in, and no browser opens again.

{
  "mcpServers": {
    "nlqdb": {
      "command": "npx",
      "args": [
        "-y",
        "@nlqdb/mcp"
      ],
      "env": {
        "NLQDB_API_KEY": "sk_mcp_REPLACE_ME"
      }
    }
  }
}

No agent handy? Try a query in the browser — an anonymous database, no sign-in, gone in 72 hours.

Questions agent builders ask

What is analytical memory for an AI agent?: A real, queryable database an agent uses to remember — typed rows it can GROUP BY, JOIN, and aggregate over in plain English, not an opaque vector blob it can only search. Recall is table stakes; analysis over memory is the part a vector store structurally can't do.
How is this different from a vector store or vector database for agent memory?: A vector store returns the top-k most similar chunks — semantic recall with no query planner, so rollups become the model doing arithmetic over search hits. nlqdb stores typed rows in Postgres, so it answers “how many times did this user ask about pricing this month?” with one GROUP BY.
Does my agent have to design a database schema first?: No. The database is created from a goal, not a schema. Your agent writes typed rows as it learns and asks questions in English; nlqdb provisions the Postgres database and infers the shape.
Is it safe to let an AI agent write SQL against its memory?: The LLM never emits SQL. It returns a typed JSON plan; nlqdb's compiler emits parameterised SQL, an AST re-parse checks it against a verb and table allow-list, and destructive writes show a diff you confirm first — and every query returns the exact SQL it ran.
Can my agent run similarity or vector recall too?: Similarity search over an embedding column (pgvector) is an opt-in slice still landing. Today the wedge is the analytical side — the GROUP BY, JOIN, and per-group aggregate queries a vector store can't run.
What does it cost, and can I use my own LLM key?: Bring your own Anthropic, OpenAI, Gemini, or OpenRouter key at 0% markup on the free tier — no card, no per-call fees. The engine, CLI, MCP server, and SDKs are source-available under FSL-1.1 and self-hostable for any non-competing use.

See the side-by-side comparisons or the manifesto.

Memory your agents can GROUP BY.