30 Lines of SQL That Beat Every AI Memory Server

Intueina · August 5, 2026 · 4 min read

Every AI memory system on the market — mem0, cognee, TencentDB Agent Memory — requires an LLM to extract meaning from stored data. They send your memories to an API, wait for a model to process them, and return results in seconds. Intueina does it in microseconds. With SQL. No API key. No model. Just geometry.

The verb-chain: what everyone else uses an LLM for.

When an AI agent needs to understand how concepts relate to each other — "how does Prima connect to memory recall?" — every other system calls an LLM. The LLM reads stored facts, infers relationships, and returns an answer. This costs time, money, and API calls. Intueina skips all of it.

Here's the entire synthesis core. 30 lines. SQLite CTE. No embeddings. No model. No network call:

WITH RECURSIVE walk(node_id, depth, names, verbs, visited) AS (
    SELECT r.to_entity, 1,
           e1.name || ' → ' || e2.name,
           r.relation_type,
           r.to_entity
    FROM relations r
    JOIN entities e1 ON e1.id = r.from_entity
    JOIN entities e2 ON e2.id = r.to_entity
    WHERE r.from_entity = ?1
    UNION ALL
    SELECT r.to_entity, w.depth + 1,
           w.names || ' → ' || e2.name,
           w.verbs || ' → ' || r.relation_type,
           w.visited || ',' || r.to_entity
    FROM relations r
    JOIN walk w ON r.from_entity = w.node_id
    JOIN entities e2 ON e2.id = r.to_entity
    WHERE w.depth < ?2
      AND w.visited NOT LIKE '%,' || r.to_entity || ',%'
  )
  SELECT node_id, depth, names, verbs FROM walk;

That's it. A recursive CTE walks the entity graph at configurable depth, building compound paths like prima → turning → recall → skill and verb chains like implements → powers → depends_on. Cycle detection prevents infinite loops. The output is a complete relationship map — in microseconds.

Live demo. Right here. Right now.

Live Verb-Chain Synthesis

Loading live data from Intueina...

The benchmark. No contest.

Metric	Intueina	mem0	cognee	TencentDB
Memory extraction	SQL CTE · Deterministic	LLM API · ~0.88s	LLM API · ~1.09s	LLM API · ~2s+
API key required	No	Yes	Yes	Yes
Offline capable	Fully	No	No	No
Relation traversal	Recursive CTE · 3 depths	LLM inference	Graph backend	LLM extraction
Latency per query	<1ms	~880ms	~1090ms	~2000ms+
Dependencies	Rust binary · Single file	Python · Docker	Python · Docker · Neo4j	Node · Docker · 3 services
Monetization pain	None · Open source	Cloud SaaS	Cloud SaaS	Cloud SaaS

<1ms

Per Query

API Keys Needed

Lines of SQL

Why this matters.

Every major memory system — from Y Combinator-backed mem0 (62k stars) to Tencent's 12k-star offering — depends on external LLM APIs for the most fundamental operation: understanding how stored information relates to itself. This is the equivalent of needing ChatGPT to read your own diary.

Intueina's verb-chain proves that relational knowledge doesn't require an LLM. It's geometry. A graph traversal. 30 lines of SQL that run on the same SQLite database every phone ships with.

When mem0 charges for API calls and TencentDB requires Docker, Intueina runs on a single Rust binary with zero external dependencies. The verb-chain is the core. The recall is the interface. Everything else — the panel, the dashboard, the cloud platform — is decoration that competitors added because they couldn't solve the core problem without an LLM.

Intueina solved it with geometry.