Introduction

MarkDB is a memory database for AI agents. It sits between your agent and the language model as an OpenAI-compatible proxy and captures every prompt, tool call and response into a structured store you can search and summarize.

Point any coding agent (Cursor, Codex or Claude Code) at the MarkDB proxy and you get durable memory, hybrid search and automatic summaries without changing how your agent works.

What you get

• Durable agent memory. Every exchange is mirrored into a structured store, organized into chats, sessions and turns.
• Hybrid search. Retrieve past work with combined vector (pgvector) and full-text (Meilisearch) search.
• Automatic enrichment. A background worker summarizes turns, sessions and whole chats so long histories stay compact and retrievable.
• One OpenAI-compatible endpoint. /v1/chat/completions, /v1/responses, /v1/messages and /v1/embeddings, dispatched natively to Anthropic, OpenAI and Gemini.
• MCP server. Expose memory to any MCP-aware client for recall and search.

How it fits together

   your agent ──▶  MarkDB proxy  ──▶  model provider (Anthropic / OpenAI / Gemini)
                       │
                       ├─▶ mirror every turn ─▶ memory store (Postgres)
                       │                              │
                       └────────────────────  enrichment worker ─▶ summaries + search index

The proxy forwards your request to the model and, in the same pass, records the exchange. A worker then enriches and indexes it in the background. Your agent sees a normal model response; MarkDB builds the memory.

Next steps

• Getting started - create an agent, mint a key, then capture and recall your first conversation.
• Operating model - how capture and recall fit together.
• Concepts - how MarkDB structures memory.
• MCP tools - the memory tools your agent uses to recall context.