Graph Hacks 2026 · Agents on connected data · Memory and coordination · Care-team brain

Some pills are treating other pills.

A side effect gets mistaken for a new illness, so a second drug is prescribed to treat it. Then a third. Domino is a team of AI agents that walk a FalkorDB medication graph, find these prescribing cascades in older adults, prove every link, and hand a pharmacist the one question worth asking.

Watch Beatrice’s review Try your own list

Beatrice, 84 · five prescribers · eighteen months

Amlodipine

blood pressure

ankle swelling

Furosemide

for the swelling

leaking urine

Oxybutynin

for the bladder

confusion

Donepezil

for the memory

9 → 4

medicines, if amlodipine is reviewed

4 → 0

anticholinergic burden

5 → 1

drugs pushing her toward a fall

Synthetic patients only. Domino raises questions for clinicians; it never tells anyone to stop a medicine.

clinically important cascades agreed by 38 clinicians in six countries

ThinkCascades, Drugs & Aging 2022

1.93×

more likely: a diuretic started after a calcium channel blocker than before it

Ireland, 533,464 adults 65+, Ann Fam Med 2025

2.5×

the precision of raw graph links (0.75 vs 0.30), at 96% recall, with zero deliberate pairs flagged

632 hidden cascades · 600 synthetic patients

401

cascades stopped at the first downstream prescription when the record is replayed with Domino at the prescriber’s desk

510 prescriptions spared · 95% of stops were real

The agent team

Three agents, one graph, and a human who approves every action.

Detective

Walks the patient’s medication subgraph: multi-hop cascade paths, feedback loops, side-effect pile-ups, the triple whammy. Records findings in the graph.

find_cascades · check_harms · recall_memory · record_findings

Pharmacist

Checks each root: timing, other reasons, published evidence, population signal. Finds safer options that don’t clash with anything else she takes, and simulates the untangle.

explain_link · find_safer_alternatives · simulate_untangle · population_signal

Scribe

Turns the evidence into review tasks for the pharmacist and prescriber, an SBAR note, and a plain-language note for the family. Everything waits for human approval.

propose_task · draft_note

Agents hand off through signed messages stored in FalkorDB and share one graph lease per patient. Clinician decisions become memory; what the team does with each task becomes a learned procedure; every question and answer is a conversation that person can pick up tomorrow; and every finding links back to the exact steps, queries and evidence behind it.

Why it needs a graph

A prescribing cascade is a path.

Drug → side effect → drug → side effect → drug. In SQL that is a pile of self-joins that breaks at every extra hop. In FalkorDB it is one variable-length pattern, constrained to the patient’s own medicines, that returns every chain in milliseconds. The same graph answers the follow-ups: what else treats this diagnosis without that side effect, which drugs converge on a fall, which pairs complete a dangerous combination.

Agents choose graph queries as tools; every answer carries the path it came from.
27 documented patterns from ThinkCascades and published studies sit in the graph as nodes.
Read tools run as GRAPH.RO_QUERY, so the database itself refuses writes.

MATCH (p:Patient {id: $pid})-[:TAKES]->(m:Drug)
WITH p, collect(m.id) AS meds
MATCH (p)-[:TAKES]->(a:Drug)-[:CAUSES]->(:Effect)-[:TREATED_BY]->(:Drug)<-[:TAKES]-(p)
WITH DISTINCT a, meds
MATCH path = (a)-[:CAUSES|TREATED_BY*2..6]->(z:Drug)
WHERE z <> a AND all(n IN nodes(path) WHERE NOT n:Drug OR n.id IN meds)
RETURN [r IN relationships(path) | [startNode(r).id, type(r), endNode(r).id]] AS hops

Returns for Beatrice · ~0.1 s

amlodipine → edema → furosemide → low potassium → potassium
amlodipine → edema → furosemide → incontinence → oxybutynin → confusion → donepezil
… 38 more paths, scored and pruned by timing, other reasons and evidence

What Domino asks of FalkorDB

Variable-length paths

Every cascade chain, up to four drugs long, inside one patient’s medicines

algo.betweenness

The drugs and effects most cascade routes pass through

allShortestPaths

How any drug reaches any harm, for the Pharmacist

Full-text indexes

Brand names with typos, and one search over everything the team wrote

A graph per clinic

Tenants kept apart at the database, chosen from the session

GRAPH.RO_QUERY