Constructive Open-Sources Postgres Memory Layer for AI Agents

The dominant pattern for giving AI agents memory is storing knowledge in markdown files. It works until it doesn't -- and for production agents handling thousands of contacts or long-running tasks, it fails predictably: the agent loads everything into the context window and hopes the model finds what matters before running out of tokens.

Constructive, the company behind open-source Postgres and JavaScript tooling with over 100 million downloads, shipped a direct answer today. Per Yahoo Finance, the firm released agentic-db as open source: a purpose-built Postgres database giving agents persistent memory, a skill and tool registry, task orchestration, runtime observability, and a full CRM and knowledge graph, all inside a single schema, installable with one command.

The timing is deliberate. Agent frameworks have matured quickly, but the infrastructure beneath them hasn't kept pace. Andrej Karpathy has publicly backed markdown-based knowledge stores as a reasonable default for agent memory. The problem isn't the format. It's the retrieval model: loading an entire history into context to answer a single query is structurally expensive, and the cost compounds as agents accumulate more state.

The retrieval problem

agentic-db doesn't eliminate markdown. It puts a query layer in front of it. The database handles search, filtering, and ranking at the database level, so agents receive only the relevant slice of knowledge rather than the full corpus. For practitioners building agents that need to reason over large structured datasets -- think CRM records, multi-session conversation history, or tool registries with hundreds of entries -- that targeted retrieval is what matters most.

Kristopher Floyd, Managing Partner at Frontier Syndicate and founder of the Bay Area Frontier Research Club -- which draws researchers from Stanford, UC Berkeley, and Google -- collaborated with Constructive early in development. His use case is concrete: hundreds of thousands of contacts, a CRM in disarray, and an agent with no practical option except dumping everything into the prompt. agentic-db was designed in part around that exact failure mode.

What's inside the schema

The release bundles components that agent developers typically build or bolt together separately: conversation history, a skill and tool registry, task orchestration, runtime observability, and a knowledge graph, all in one Postgres schema. Having these in a single database means agents can run transactional queries across memory types, rather than stitching results from separate stores.

This is a different architectural bet than the vector-database-first approach that dominated artificial intelligence infrastructure discussions through 2023 and 2024. Pure vector search excels at semantic lookup but struggles with structured filtering, relational queries, and transactional writes. Postgres handles all of these natively. Hybrid retrieval -- combining vector similarity with structured SQL -- is increasingly what practitioners reach for in production, and as llm-stats.com documents, the pace of new agent-capable model releases only raises the stakes for getting memory infrastructure right.

Constructive is releasing this under an open-source license, meaning teams can self-host and inspect the full schema. That matters for enterprise adoption, where routing all agent memory through a third-party API is often a compliance non-starter.

Agent infrastructure goes wider

OpenAI recently shipped Symphony, which Gizmodo described as a framework connecting coding agents to project management boards like Linear -- turning every open task into an autonomous agent run. These tools operate at different levels of the stack. Symphony orchestrates agent execution; agentic-db governs what agents know and remember. They are complementary, not competing.

Beyond individual companies, investment in agent infrastructure is moving into national science programs. South Korea recently partnered with Google DeepMind on its K-Moonshot initiative, as UPI reported, targeting AI-driven breakthroughs in life sciences, climate research, and robotics. Serious artificial intelligence investment now extends to the infrastructure layer, not just the models sitting on top of it.

Agents that forget nothing and retrieve precisely may be closer than they appeared six months ago. Whether practitioners will actually migrate memory backends -- away from the markdown-plus-context-window pattern that currently dominates -- is the more interesting question.

FAQ

What is agentic-db?
agentic-db is an open-source Postgres database architecture from Constructive that gives AI agents persistent memory, conversation history, a skill and tool registry, task orchestration, runtime observability, and a knowledge graph, packaged in a single installable schema.

How does agentic-db differ from a vector database?
Vector databases are optimized for semantic similarity search but struggle with structured filtering and transactional writes. agentic-db uses Postgres natively, enabling hybrid retrieval that combines vector similarity with standard SQL queries and relational operations.

What is context window bloat in AI agents?
Context window bloat happens when an agent loads its entire knowledge base or conversation history into the model's context window to answer a query, consuming tokens and risking truncation. agentic-db addresses this by retrieving only the relevant data at query time.

Is agentic-db free to use?
Constructive released agentic-db as open source, allowing teams to self-host the full schema without routing data through a third-party API.