Memobase – Unified Memory Layer for AI Agents

TL;DR

TL;DR: Memobase is a persistent memory server for AI agents that lets Claude, Cursor, and ChatGPT share context across sessions and tools — with a fully self-hosted option that keeps your data in a local SQLite database.

What Is Memobase?

Every AI agent session starts from a blank slate. Intelligence built up over hours in one tool is invisible to another. Memobase (memobase.ai) is a shared memory layer that sits behind your AI tools, giving them a persistent, queryable graph of what you have been working on.

The core claims from the HN launch thread:

“There is no standard protocol for AI memory. You cannot say ‘here is my MCP server, use it for memory in every session.’ Each platform locks you into its own memory store.”

Memobase attempts to solve this by exposing a Model Context Protocol (MCP) server and OpenAI Actions interface that any compatible agent can hook into.

How It Works

Memobase combines three technologies to maintain agent memory:

Knowledge Graph — Every memory entry is parsed for entities (people, projects, concepts) and their relationships. Agents can navigate context via association rather than keyword matching.

Hybrid Search — Combines semantic vector search with BM25 keyword search. The docs describe it as “understanding meaning” plus “exact term matching” to cover both fuzzy and precise queries.

Memory Consolidation — Background workers deduplicate, merge, and update memories as new information arrives, keeping the knowledge base current rather than just appending logs.

Setup Workflow

Step 1: Install the CLI

npm install -g memobase-cli

Requires Node.js. The CLI handles authentication and agent configuration.

Step 2: Start in Local-First Mode

Memobase supports two modes. For full privacy, use local mode:

mb start

This starts a local daemon and MCP server. All memories are stored in ~/.memobase/local.db — a local SQLite database. No data is sent to Memobase servers.

Step 3: Inject into Your AI Tool

mb inject claude-code

This injects Memobase into the specified agent. Supported targets include Claude Code, ChatGPT, and Cursor.

Self-Hosted vs. Cloud

The local SQLite mode stores everything on-disk and never phones home. However, entity extraction and embeddings still require an AI engine. You can provide your own OPENAI_API_KEY to keep processing private to your own account.

The managed cloud mode syncs across devices and handles embedding/computation on Memobase infrastructure.

Source and Accuracy Notes

• Project page: memobase.ai
• Documentation: memobase.ai/docs
• HN launch thread: news.ycombinator.com/item?id=47237920
• License: Not yet verified from a LICENSE file
• Source last checked: 2026-06-25

FAQ

Q: Does Memobase require an API key? A: Local mode still requires an OPENAI_API_KEY for entity extraction and embeddings, but your memory data stays local. The cloud mode uses Memobase infrastructure for both storage and processing.

Q: Which AI tools are supported? A: The docs mention Claude Code, ChatGPT, and Cursor as injection targets. The MCP server interface means any MCP-compatible agent can connect.

Q: Is there a free tier? A: The website mentions a free tier with 1000 memory entries. Pricing details beyond that were not visible in the docs at time of writing.

Q: How does memory consolidation work? A: According to the docs, background workers deduplicate, merge, and update memories as new information arrives — rather than appending a flat log of every interaction.

Conclusion

Memobase is a concrete attempt to solve the AI memory silo problem by building a shared knowledge graph accessible via MCP. The local-first mode with SQLite storage is the standout feature for privacy-conscious users — your conversation context never leaves your machine unless you explicitly opt into cloud sync. Worth evaluating if you regularly switch between multiple AI agents and are tired of re-explaining your project context each time.