Constructive launches agentic-db for AI agents

Constructive has released agentic-db, an open-source Postgres database for AI agents that provides persistent memory within a single database.

The product targets developers building AI agents that need to store and retrieve conversation history, structured knowledge and operational records without relying on file-based memory systems. Released under the MIT licence, it can be installed as a single Postgres schema through pgpm, which Constructive describes as a package manager for PostgreSQL modules.

Constructive argues that many agent systems still rely on markdown files to store knowledge and history, forcing models to load large volumes of text into prompts to answer simple questions. Agentic-db is intended to move that work into the database layer by enabling query-based retrieval across stored data.

The release adds a broad set of functions within Postgres, including long-term memory storage, conversation and tool tracking, registries for skills and prompts, rules and behavioural policies, task orchestration, runtime observability and a world model covering records such as contacts, companies, deals, emails, calendars, projects and goals.

It also supports multiple retrieval methods across text-bearing tables, including semantic vector search via pgvector, BM25 ranking, weighted full-text search via tsvector, trigram fuzzy matching, and PostGIS spatial queries.

An automatic embedding pipeline updates embeddings via Postgres triggers and Ollama, rather than via cron jobs or external services. The database is intended for local and single-developer use.

Dan Lynch, Founder and Chief Executive Officer of Constructive, outlined the company's view of where agent systems should begin.

"We built agentic-db because we believe agent-first infrastructure starts with the database," said Lynch. "Agents need to remember, search, and reason over structured knowledge - not re-read a pile of markdown files every time they wake up. With Constructive, we're making Postgres the foundation of that infrastructure, and open-sourcing it so every developer building agents can start with a real brain instead of a scratch pad."

Database tooling

The software is built on Constructive's own infrastructure and generates tools directly from the database schema. These include Agent Skills, instruction files to help AI assistants work with the database, a command-line interface, and a type-safe software development kit.

Constructive said the skills are available for Claude, Claude Code, Cursor, Devin, Copilot, Windsurf and Codex, among others. The command-line interface includes create, read, update, delete, search and administrative commands for every table, while the SDK uses a Prisma-like interface.

Lynch described the generated tools as a central part of the product's design. "Every database Constructive generates is agent-first - the Skills, the CLI, and the SDK all come from the schema. Agentic-db is the first open-source proof of that. We're giving agents a real database with the tooling already wired up, so developers can focus on what their agents do instead of how they connect."

Constructive said it has recorded more than 100 million open-source downloads across its software. Its technology is also used by platforms including Supabase, Neon and Gel Data.

Early users

Kristopher Floyd, Managing Partner at Frontier Syndicate, said he worked with Constructive during the database's early development. He described the project as a response to a practical problem involving large volumes of contact data and poor searchability.

"I have hundreds of thousands of contacts, and my CRM is a disaster. My agent's only option was to dump everything it knew into the prompt and hope for the best," said Floyd. "agentic-db gives agents vector search, full-text search, and auto-embeddings out of the box - so they can actually search what they know instead of brute-forcing it every time. I see Constructive becoming a core piece of infrastructure for the agentic internet."

Another early user, Michael Shawn Lockwood, said he deployed the system to a bare-metal Debian virtual machine. Lockwood, a database architect and engineer, highlighted the breadth of functions included in the schema.

"This is one of those rare moments where you stop and think: someone spent years building this," said Lockwood. "What looks like 'just a database' is actually a full foundation layer for agentic systems - long-term memory, conversations, tool execution tracking, task orchestration, hybrid retrieval - and it's all built on top of PostgreSQL. I have a strong feeling this becomes a go-to foundation layer for real-world systems over the next few years," said Lockwood