Comparison
nlqdb vs Letta
Pick Letta if you want a stateful agent runtime that manages its own memory like an OS — self-editing core blocks plus a searchable archive. Pick nlqdb if your agent also needs to aggregate that memory: GROUP BY, JOIN, and HAVING over structured rows it provisions and migrates itself in plain English.
Persona this comparison serves: P2 agent builder. Letta's positioning: Open-source (Apache-2.0) agent runtime with persistent memory built in — OS-style tiers: core blocks the agent self-edits, searchable recall, and an archival store queried via tool calls.
When to choose nlqdb
- Your agent must aggregate its memory (GROUP BY, JOIN, HAVING), not just recall facts.
- You store structured rows the agent later reports over ('deals per stage this quarter').
- The schema should evolve as the agent learns ('add a `priority` field') via English.
- You want a data layer that drops into any runtime, not a runtime itself.
When to choose Letta
- You want a full stateful agent runtime, not just a store the agent queries.
- OS-style self-editing memory (core / recall / archival tiers) is the model you need.
- You want an Apache-2.0 core you can self-host with no source-available terms.
- Memory is prose the agent edits and searches — relevant-fact retrieval over typed SQL.
In your HTML
The relational aggregation Letta's memory tiers can't run — recall and archival search return relevant entries, not a GROUP BY / AVG result; nlqdb answers it as SQL over the agent's own memory.
>
average deal size per stage across everything the agent logged this quarter <nlq-data goal="average deal size per stage across everything the agent logged this quarter"></nlq-data>Feature parity, honest
| Feature | nlqdb | Letta | Note |
|---|---|---|---|
| Owns the database (provisions + migrates) | |||
| Natural-language → SQL | Letta's agent reads and writes memory through tool calls (core-block edits, recall/archival search); it has no NL→SQL compiler over a relational store. | ||
| Aggregations + reporting queries (GROUP BY / JOIN / HAVING over memory) | Letta can recall 'Alice has a $50k deal' but can't answer 'average deal size per stage' — there is no relational query layer over the memory tiers. | ||
| OS-style self-editing memory tiers (core / recall / archival) | Editable core blocks always in context, plus searchable recall and archival, are Letta's core primitive; nlqdb stores typed rows in Postgres tables, not self-edited memory blocks. | ||
| Stateful agent runtime (the agent loop itself) | Letta runs the agent; nlqdb is the data layer that any runtime — Letta included — can provision and query. They compose: Letta the runtime, nlqdb the analytical store. | ||
| Auto-migration via NL ('add a `priority` field') | |||
| MCP server (agent-callable) | Letta supports MCP integration (it can consume MCP tools and exposes its agents over MCP); nlqdb's MCP exposes `nlqdb_query` / `nlqdb_list_databases` / `nlqdb_describe`, and `nlqdb_query` materialises Postgres on first reference. | ||
| Vector / semantic recall over archived memory | Letta's archival tier indexes memory with embeddings for semantic search; nlqdb answers with structured SQL and ships no embedding-based recall today. | ||
| Open source / self-hostable | Letta is Apache-2.0 (self-host or hosted). nlqdb is source-available on FSL 1.1-ALv2, auto-converting to Apache 2.0 after two years. |
shipped · partial · not shipped
Questions buyers ask
- Can I use Letta as the agent runtime and nlqdb as the store it queries?
- Yes — they compose. Letta runs the stateful agent and manages its OS-style memory tiers; nlqdb is the analytical store the agent provisions and queries with SQL. Use Letta for the agent loop and self-edited recall, nlqdb for 'average deal size per stage across everything the agent logged'.
- Does nlqdb give an agent self-editing memory tiers like Letta's core / recall / archival?
- No. nlqdb is Postgres-first — typed rows in tables, not self-edited prose blocks in a context window. If OS-style memory the agent rewrites itself is core to your design, Letta (out of the MemGPT paper) is the right shape; nlqdb's contract is relational SQL over the rows the agent provisions.
- Letta supports MCP — how is nlqdb's MCP different?
- Letta's MCP support lets its agent consume external tools and exposes Letta agents to MCP clients. nlqdb's MCP server (`mcp.nlqdb.com`) exposes `nlqdb_query`, `nlqdb_list_databases`, and `nlqdb_describe`; `nlqdb_query` materialises Postgres plus schema on first reference and runs aggregating SQL, so the agent reports over its memory instead of only retrieving from it.
- Can Letta aggregate an agent's memory the way nlqdb can?
- Not relationally. Letta's archival memory is searched semantically and recall is conversation history — both return relevant entries, not a GROUP BY / JOIN / HAVING result set. nlqdb compiles the English question to SQL and runs it over typed rows, so the agent gets a real aggregation, not a ranked list of matches.
- Can my AI agent provision its own store with Letta the way it can with nlqdb?
- Letta provisions an agent with memory tiers, but the archival store is backed by a database you configure — it doesn't stand up an isolated relational database the agent migrates. nlqdb's MCP `nlqdb_query` materialises a tenant-scoped Postgres plus schema on first reference, so a Claude / Cursor / Cline agent provisions and migrates its data layer end-to-end without a human in the loop.
Try nlqdb in 30 seconds
No sign-in. The anonymous database lasts 72 hours; adopt it with one click if you keep it.
Start with a goal →