Services

What we build on your Databricks.

Four practice areas, three engagement shapes, one architecture pattern that respects your existing Lakehouse. Engagements delivered through Next Link Labs.

Beyond thread-level checkpointing

What Databricks’ guidance leaves to the implementor.

Databricks’ website walks through thread-level checkpointing with Lakebase, the short-term memory layer for AI agents. It also notes, explicitly, that longer-lived memory “introduces additional design considerations around scope, retention, and governance.” That’s where we live.

  • Long-term memory across threads and users: accumulated knowledge with deliberate scope.
  • Multi-tenant RLS on memory tables: separation that survives audit.
  • Retention & governance: what stays, what expires, who decides.
  • Memory regression evals: automated checks that memory keeps helping, not hurting.
  • Observability: what each agent remembered, retrieved, and why.
  • Custom retrieval strategies: beyond off-the-shelf vector search.
Practice areas
01
Memory

Memory layer implementation

Episodic, semantic, and working memory backed by Unity Catalog tables and Vector Search. Built on our open-source `lakehouse-memory` foundation, with the production extensions (compaction at scale, multi-tenant RLS, regression evals, observability, custom retrieval) that aren’t in OSS.

02
RAG

RAG production-ization

Move RAG from notebook demos to versioned pipelines: ingestion, embedding refresh, retrieval evals, drift monitoring, MLflow tracking.

03
Agents

Agent platforms on Databricks

Multi-step agents with state, tool use, and observable runs. Built on Mosaic AI Agent Framework where it fits, custom where it doesn’t.

04
Claude

Claude integration

Wire Claude into the reasoning, planning, and eval layers without a vendor rewrite. Cost-aware routing across model providers.

Engagement models

Assessment

Fixed-fee · 2 weeks

Workspace audit, architecture document, prioritized backlog. Decision-ready output.

Implementation

T&M · 6–12 weeks

We build the memory layer + your highest-priority use case end to end, in your workspace, with your team.

Advisory retainer

Monthly · Ongoing

Architecture review, code review, sounding-board for your AI platform team. Cap on hours, no lock-in.

Reference architecture

Memory at the center.

Same diagram as the homepage, expanded. The memory layer reads and writes Unity Catalog tables. Claude handles reasoning. MLflow captures evals. Your existing access control governs everything.

Agent Memory Layer UC Tables Claude Eval MLflow
Stack we work with
  • Unity Catalog
  • Lakebase
  • MLflow
  • Mosaic AI Agent Framework
  • LangGraph
  • Vector Search
  • DBSQL
  • Delta Live Tables
  • Claude (Anthropic)
  • OpenAI / OSS LLMs

Want to see how this fits your workspace?

A 30-minute call is enough to know whether memory-aware AI on your Databricks is worth pursuing.

Book a call →