Python memory helpers
connection.memory wraps memory writes, searches, and context assembly. connection.cache wraps exact and semantic prompt cache store/lookup.
Python memory helpers
connection.memory wraps memory writes, searches, and context assembly. connection.cache wraps exact and semantic prompt cache store/lookup.
LangGraph-style flows
Example retrieve/call/remember flows show how to assemble context, check cache, call a provider on miss, and write memory back.
Provider adapters
Optional example adapters cover OpenAI Responses and Anthropic Messages. They lazy-import SDKs and keep model names in application config.
AgentOps telemetry
Example schemas model agent_runs, retrieval_events, cache_events, llm_calls, llm_errors, tool_calls, context_traces, and context_replay_events as ordinary time-series tables.
OpenTelemetry mapping
Example mapping converts OTLP-style GenAI spans into ZeptoDB AgentOps tables for LLM calls, cache hits, tool calls, and errors.
ZeptoDB does not call LLM or embedding providers from the server. Applications own providers and credentials; ZeptoDB stores memory, filters candidates, assembles context, and checks cache.
Arrow IPC ingest
POST /insert/arrow accepts Apache Arrow IPC RecordBatchStream bytes and writes through the table-aware tick ingest path without per-row SQL.
MessagePack ingest
POST /insert/msgpack accepts a dependency-light MessagePack map of column arrays for embedded and Telegraf-style batched clients.
Telegraf output
zepto-telegraf-output plugs into Telegraf outputs.execd, reads Influx line protocol from stdin, and writes ZeptoDB batches over HTTP.
Kafka
Built-in Kafka consumer with batch ingestion. Up to 3.2M events/sec from Kafka topics. Useful for operational event streams and AgentOps pipelines.
Kafka connector behavior should be validated in your deployment shape before relying on it.
FIX 4.4
Native FIX parser for market-data and order feeds. Trading agents can pair tick data with strategy memory and decision replay.
ITCH
NASDAQ ITCH handler for L3 order-book data. Lossless, nanosecond-timestamped, and ready for temporal market-memory workflows.
Binance WebSocket
Real-time crypto market data ingestion for agents watching multi-exchange state.
AWS Kinesis
Optional KinesisConsumer for AWS-native streams, reusing Kafka/MQTT tick decoders with table-aware local and cluster routing.
ROS 2
Optional Ros2Consumer bridge for Physical AI: scalar topics, standard message profiles, typed profile tables, cluster forwarding, and rosbag2 import/replay.
ROS 2 Setup → · Edge Deployment → · C++ Reference →
ROS 2 connector support requires
-DZEPTO_USE_ROS2=ON; IoT connectors should be validated against your edge and operations requirements.
| Language | Type | Typical use | Link |
|---|---|---|---|
| Python | Zero-copy + memory/cache helpers | Agent loops, ML pipelines, notebooks | Python Reference → |
| C++ | Native Pipeline API + connector helpers | Embedded, low-latency, edge, robot-lab, and streaming deployments | C++ Reference → |
| HTTP | REST API + binary ingest | SQL, /api/ai/*, Arrow IPC ingest, and MessagePack ingest | HTTP Reference → |
| Arrow Flight | gRPC streaming | Batch data transfer and distributed training | Flight Reference → |
Prometheus
Built-in /metrics endpoint. Scrape ingestion rates, query latency, memory usage, cluster health, Agent Memory counts, cache counts, evictions, and capacity limits.
Grafana
Query ZeptoDB directly from Grafana via HTTP API. Build dashboards over operational telemetry, AgentOps events, and memory/cache metrics.
| Integration | Status | Description |
|---|---|---|
| Agent Memory sidecar | ✓ | records.bin + vectors.bin snapshots, WAL replay, snapshot metrics, ANN footprint, and sidecar byte gauges |
| Amazon S3 | ✓ | Parquet HDB for historical time-series data |
| Google Cloud Storage | ✓ | Parquet HDB |
| Local NFS | ✓ | On-premise Parquet HDB for air-gapped and edge |
| Docker | ✓ | Single-command deployment and Agent Memory pilots |
| Kubernetes | ✓ | Helm chart with StatefulSet for time-series clusters |
| ARM Graviton | ✓ | aarch64 native for cost-efficient edge and cloud |
| Provider | Protocol | Status |
|---|---|---|
| Okta | OIDC | ✓ |
| Auth0 | OIDC | ✓ |
| Azure AD / Entra ID | OIDC | ✓ |
| Google Workspace | OIDC | ✓ |
| API Keys | Native | ✓ |
| JWT | Bearer token | ✓ |
| Integration | Status | Description |
|---|---|---|
| OPC-UA | Available / expanding | PLC / SCADA connector with scalar, array, string, structured, Historical Access, Alarms & Conditions, browse CLI, and server-mode surfaces |
| MQTT | Available / expanding | Lightweight IoT sensor and robot telemetry ingestion |
| ROS 2 scalar topics | Available | Live scalar std_msgs topic subscriptions plus rosbag2 import/replay |
| ROS 2 standard profiles | Available | IMU, JointState, Odometry, TFMessage, and LaserScan scalar mapping |
| ROS 2 typed profiles | Available | Schema-aware wide tables for standard profiles, with typed-row cluster forwarding |
| AutomationML / PackML | Under evaluation | Standardized plant and machine data models |
| CAN / Vehicle Bus | Under evaluation | Automotive telemetry ingestion at the edge |
| Integration | Status | Description |
|---|---|---|
| AWS Kinesis | Available | Optional AWS-native stream consumer with JSON/BINARY/JSON_HUMAN tick decoders |
| Telegraf output | Available | outputs.execd writer for Telegraf input ecosystems |
| MessagePack ingest | Available | Dependency-light HTTP binary batch ingest at /insert/msgpack |
| Arrow IPC ingest | Available | Arrow RecordBatchStream HTTP ingest at /insert/arrow |
| Apache Pulsar | Planned | Streaming alternative to Kafka |
| Kafka Connect Sink | Planned | Standard Kafka connector plugin |
| CDC (Debezium) | Planned | PostgreSQL / MySQL change capture |
| JDBC / ODBC | Planned | Tableau, Excel, BI tools |
| DuckDB | Planned | Arrow zero-copy JOIN delegation |
Interested in an integration, or want to compare implementation notes? Open an issue on GitHub or start a GitHub Discussion.