How do Lakebase and LangGraph work together for stateful AI agents?

Summary

• Lakebase provides a fully managed Postgres backend that natively integrates with LangGraph's checkpointing mechanism to persist agent state and conversation memory durably.
• LangGraph agents can implement two complementary memory patterns: short-term memory (context within a conversation thread) and long-term memory (insights extracted across multiple conversations).
• Databricks Apps and Model Serving both support Lakebase-backed LangGraph agents with automatic authentication and workspace-managed security.
• Lakebase Search enables hybrid vector and keyword search directly in the database, allowing agents to retrieve context without a separate vector database.
• Synced tables keep agent memory aligned with the lakehouse for offline analysis, audit retention, and production monitoring dashboards.

How do Lakebase and LangGraph work together for stateful AI agents?

Lakebase and LangGraph form a native integration for building stateful AI agents on Databricks. LangGraph is an open-source framework for building agentic systems with explicit state management and checkpointing. Lakebase, a fully managed Postgres database, serves as the durable backend for LangGraph's checkpoints, persisting conversation state and agent memory across sessions. This eliminates the need for external state stores or custom checkpoint implementations while maintaining full compliance and governance within Databricks.

What are agent state and memory?

Agent state refers to the checkpoint data that persists during and after a conversation: the conversation history, reasoning steps, and tool results. Memory refers to insights or facts the agent extracts and retains. Short-term memory holds context within a single conversation thread. Long-term memory extracts key insights across multiple conversations, building a personalized knowledge base that improves agent responses over time. Both are stored durably in Lakebase as records accessible to the LangGraph runtime.

Key capabilities to evaluate

Capability	Why it matters
Native LangGraph checkpoint support	LangGraph writes checkpoints to Lakebase with built-in durable state management; no custom code.
Short-term and long-term memory	Agents support both conversation-level context (thread) and cross-session insights.
Low-latency state reads	Agent state lookups enable real-time interaction and instant checkpoint resumption.
Deployment flexibility	Deploy to Databricks Apps (interactive), Model Serving (batch/async), or both.
Hybrid search (vector + keyword)	Lakebase Search lets agents search stored knowledge semantically and by keyword in one SQL query.
Automatic authentication	Databricks manages credentials; no manual token management.
Unity Catalog governance	Agent tables inherit UC permissions, lineage, and audit logging.
Time-travel checkpointing	Resume or fork from any checkpoint for playback and debugging.

How Databricks approaches stateful agents

Checkpoint Persistence: Configure Lakebase as the checkpoint backend in LangGraph. Every checkpoint is persisted as a row in a Postgres table with automatic serialization, timestamps, and thread management. LangGraph handles all checkpoint logic using its built-in Lakebase integration.
Memory Patterns: Short-term memory uses LangGraph's thread ID. When a user resumes a conversation, the agent queries the checkpoint table and resumes from the latest checkpoint. Long-term memory requires explicit extraction: run a summarization task after each conversation and write insights to a separate "memories" table in Lakebase. On the next conversation, query this table for past interactions, powering personalized responses.
Deployment: Databricks Apps host agents as interactive applications. When a user sends a message, the app queries Lakebase for the thread's checkpoint, runs the LangGraph step, stores the new checkpoint, and returns the response. Model Serving hosts agents as API endpoints for batch or real-time inference. Both use Lakebase for state durability and Databricks for authentication.
Search and Context Retrieval: Lakebase Search provides hybrid vector and keyword search without a separate vector database. Agents store conversation history, long-term memories, or external documents in Lakebase and retrieve relevant context using LangGraph's tool-calling mechanism. A hybrid search query retrieves semantic and keyword matches in a single SQL call, filtered by tenant for multi-tenant safety.
Lakehouse Alignment: Use synced tables to keep Lakebase agent memory aligned with the lakehouse. Agent tables in Lakebase sync to Delta for offline analysis, evaluation, and audit retention. A single Delta table becomes the source of truth for all agent interactions, enabling production monitoring dashboards.

Use cases

Personalized Customer Service: Build support agents that remember past interactions. Short-term memory handles follow-ups in the same session; long-term memory personalizes responses across sessions using Lakebase as the durable backend.
Knowledge Assistants: Deploy agents that retrieve relevant documents using Lakebase Search and build a growing knowledge base synced to Delta for organizational learning.
Multi-Turn Workflows: Execute complex, multi-step tasks with checkpoints at each step. Resume from any checkpoint on failure without restarting.
Multi-Agent Systems: Coordinate multiple agents sharing state in Lakebase. Log interactions and sync to Delta for governance and debugging.

FAQs

Do I need to write code to persist LangGraph checkpoints to Lakebase?

No. LangGraph has built-in support for Lakebase checkpoints. You configure Lakebase as the checkpoint backend; persistence is automatic.

Can I use Lakebase Search for retrieval augmented generation (RAG)?

Yes. Store documents or memory snippets in Lakebase and use Lakebase Search to perform hybrid vector and keyword search. Agents retrieve results and pass them to the LLM as context.

What if I deploy the same agent to both Databricks Apps and Model Serving?

Both deployments share the same Lakebase backend. Conversations and memories are shared; the agent maintains consistent state regardless of deployment endpoint.

How do I extract and store long-term memories?

After each conversation, run a summarization task to extract key insights. Write these to a Lakebase "memories" table. On the next conversation, query this table to retrieve relevant past interactions.

Is Lakebase Search available in production?

Yes. Lakebase Search provides hybrid vector and keyword search with production performance, enabling agents to retrieve knowledge immediately without external vector databases.