Build AI agents with SynapCores
One engine for SQL, vectors, graph, and LLM calls — so your agent's memory, retrieval, and reasoning live next to your data instead of across five services. Import your data, write your first agent, and ship it to production. All on this page.
Memory that persists
MEMORY_STORE and MEMORY_RECALL give every agent durable, semantic recall across sessions — no Redis, no separate memory service.
One engine, zero sync
Rows, vectors, and graph share one storage engine. No dual-write reconciliation job between a graph DB and a vector DB.
Explainable by default
Every AGENT_RUN call and every durable-agent run lands in a tamper-evident, hash-chained audit table you can query.
Production from day one
CREATE AGENT ships a scheduled or event-driven agent as a database object — not a cron script duct-taped to a queue.
How an agent query flows
No orchestration layer to stand up. Your app talks to one gateway; the agent's tools talk to the same storage engine your rows already live in.
Your app
sends a question or task
AGENT_RUN
LLM plans, calls execute_query / semantic_search / graph_query / rag_search
SynapCores
rows + vectors + graph, one storage engine, one transaction
The playbook: import data → first agent → production
Seven steps, ~2 hours end to end, every snippet copy-pasteable against the free Community Edition. Same progression graph database vendors teach — except the graph, the vectors, and the agent runtime are one download.
A table can hold a VECTOR column, a Cypher query can traverse the same rows, and an LLM agent can call both — all on one connection, no ETL between systems.
-- One engine. One connection. Four primitives an agent needs:
CREATE TABLE support_tickets (
id INT PRIMARY KEY,
subject TEXT,
body TEXT,
embedding VECTOR(384) -- vector search lives on the row
);
-- Cypher graph traversal, same connection, same data:
MATCH (t:Ticket)-[:FILED_BY]->(c:Customer) RETURN t, c LIMIT 5;
-- An LLM agent that can reason over both, in one call:
SELECT AGENT_RUN('aidb-assistant', 'Summarize open tickets for customer 42');Weighing Neo4j + a vector DB against a unified engine?
Step 6 above is the short version. The full architecture breakdown — what Neo4j does better, where the dual-store sync tax shows up, and when a single engine wins — is in GraphRAG vs Neo4j: Picking the Right Tool for the Right Job.
Where to go next
More agent recipes
Multi-agent shared memory, semantic tool routing, RAG with citations, and more — every recipe is runnable Cypher or SQL.
Browse agent recipesFull API reference
Every endpoint used above, plus auth, transactions, and NL→SQL, shown in cURL, Node, Python, and PHP.
Open developer docsWant us to build it with you?
The Agent Memory JumpStart is a founder-led sprint that wires SynapCores into your real workflow.
Apply for JumpStartRun the tutorial above on your own machine
Single binary, no GPU required, Community Edition is free.