NoSQL vs. NewSQL [2026] Cheat Sheet for Agent Swarms

How to decide fast

Choose NoSQL when:

• Each agent mostly owns its own document, partition, or append-only event stream.
• You optimize for predictable low-latency reads and writes on known access patterns.
• You need flexible schemas for tool outputs, traces, prompt artifacts, or episodic memory.
• You plan to drive downstream workers from built-in change streams or table streams.
• Your strongest consistency requirement is usually per item, not across many entities.

Choose NewSQL when:

• Multiple agents compete to claim, mutate, or reconcile the same rows at once.
• You need SQL joins for planners, evaluators, billing, audit, or tenant-wide analytics.
• Idempotency, uniqueness, and state machine transitions must be enforced in the database.
• You want one operational store for OLTP correctness and standard relational tooling.
• Regional scale matters, but you do not want to give up transactional semantics.

Live command filter

Use the filter to narrow examples by engine, concern, or workflow. This section is intentionally broad: bootstrap, consistency, CDC, and schema change are the four knobs that matter most for swarm backends.

mongosh 'mongodb://localhost:27017'
use swarm
db.createCollection('runs')

aws dynamodb create-table \
  --table-name SwarmTasks \
  --attribute-definitions AttributeName=AgentId,AttributeType=S AttributeName=TaskId,AttributeType=S \
  --key-schema AttributeName=AgentId,KeyType=HASH AttributeName=TaskId,KeyType=RANGE \
  --billing-mode PAY_PER_REQUEST

cockroach demo
cockroach sql --execute='CREATE DATABASE swarm;'

gcloud spanner instances create my-instance-id \
  --config=regional-us-east1 \
  --description=my-instance-display-name \
  --nodes=3

const stream = db.runs.watch()

aws dynamodb update-table \
  --table-name SwarmTasks \
  --stream-specification StreamEnabled=true,StreamViewType=NEW_AND_OLD_IMAGES

CONSISTENCY LOCAL_QUORUM
SERIAL CONSISTENCY LOCAL_SERIAL

gcloud spanner databases ddl update my-database-id \
  --instance=my-instance-id \
  --ddl='ALTER TABLE test_table ADD COLUMN a INT64'

Keyboard shortcuts

The table below matches the live filter script embedded in this page.

Commands by purpose

Bootstrap

• DynamoDB: use --billing-mode PAY_PER_REQUEST first unless you already know traffic shape.
• CockroachDB: use cockroach demo for quick local modeling of transactional agent flows.
• Spanner: provision the instance before worrying about schema; replication geography is an early architectural decision.
• MongoDB: start by modeling one run, task, or memory object as one document boundary.

Consistency and transactions

• Cassandra: consistency level is a first-class knob, so encode it in application policy rather than scattering it ad hoc.
• CockroachDB: default SERIALIZABLE isolation is a strong fit for task-claiming, quotas, and shared budget counters.
• MongoDB: use multi-document transactions selectively when a single-document model is no longer enough.
• Spanner: favor it when correctness has to hold across rows, services, and regions.

CDC and agent fan-out

• DynamoDB Streams are a straightforward path to event-driven workers, retries, and audit reconstruction.
• MongoDB change streams work well for reactive agents that watch collection, database, or deployment-level changes.
• Inference: if your swarm architecture already depends on outbox-style eventing, built-in streams reduce custom plumbing more than they reduce latency.

Configuration patterns

Pattern 1: split operational state from long-term memory

• Put strict workflow state in NewSQL: task ownership, retries, locks, quotas, tenant billing, and audit rows.
• Put flexible memory in NoSQL: tool traces, prompt payloads, summaries, embeddings metadata, and agent scratchpads.
• Do not force vector-adjacent memory into a relational core if its shape changes every week.

Pattern 2: shape partitions around conflict domains

• Partition by tenant_id when most contention is intra-tenant.
• Partition by conversation_id or run_id when each swarm execution is mostly isolated.
• Avoid partition keys that make all schedulers hammer the same hot row or hot hash key.

Pattern 3: treat CDC as product surface, not plumbing

• Define which writes emit downstream work and which writes are purely internal.
• Make every emitted event idempotent and replay-safe.
• If you need to share sample payloads outside production, scrub them first with the Data Masking Tool.

Advanced usage

Use a two-tier write path for swarm correctness

• Tier 1 stores the authoritative task state.
• Tier 2 stores derived memory, search material, and downstream event consumers.
• This reduces the temptation to make a single database do transactional state, semantic retrieval, event fan-out, and analytics equally well.

Decide where retries terminate

• In NewSQL, let unique constraints, transaction retries, and state transitions absorb duplicate work.
• In NoSQL, push idempotency keys and compare-and-set semantics into the application contract.
• Document which agent retries are safe, and which must dead-letter after semantic conflict.

Vector and RAG considerations

• MongoDB Vector Search supports ANN and ENN search and is a practical fit when you want document state and retrieval in one platform.
• Inference: for most swarms, retrieval memory and authoritative workflow state should remain loosely coupled even if the same vendor can host both.
• Keep embeddings, summaries, and raw execution artifacts on a cheaper update path than consensus-heavy workflow rows.

Verified vendor docs

• DynamoDB Streams
• DynamoDB transactions
• MongoDB transactions
• MongoDB change streams
• MongoDB Vector Search
• CockroachDB SERIALIZABLE isolation
• Cloud Spanner documentation
• Cassandra consistency levels