Early access — memory layer for AI chats

Stop re-explaining yourself to AI.

MemoryOS turns every Claude, ChatGPT, and Gemini session into a dot in a connected memory graph — and serves the right context to your next conversation. Local-first. Private by default.

Works with Claude ChatGPT Gemini — one memory across all three

Your context is scattered across vendors.

Every new chat starts from zero. You've explained your stack to Claude, your project to ChatGPT, your travel plans to Gemini — and tomorrow you'll explain them all again. Each assistant keeps its own memory, locked inside its own walls. None of them talk to each other, and none of it belongs to you.

MemoryOS sits underneath all of them. One memory, one graph, every session connected — stored on your machine, served wherever you chat next.


How it works

01

Capture

When a session ends, it's saved as a dot: a compact summary of what was decided, with its source and tags.

02

Connect

New dots link themselves to related ones through shared tags and topics. No manual filing — the graph assembles itself as you work.

03

Recall

Start a new session and ask. MemoryOS finds the matching dots and their connections, and serves a compact, token-budgeted context block.

What it does

A graph, not a pile

Sessions connect through shared topics into a navigable web — like a graph view for your AI history. Hubs emerge around the things you actually work on.

Cross-LLM by design

One memory across Claude, ChatGPT, and Gemini. Built-in memories are vendor silos — this one travels with you.

Local-first, private by default

Everything lives in a single SQLite file on your machine. No cloud, no account, no telemetry. Delete the file and it's gone — it's that much yours.

Self-assembling

Dots auto-link on save when they genuinely share content. Connected, never a hairball — there's a threshold, not magic.

Recall that serves, not dumps

Hybrid retrieval — full-text search fused with semantic similarity, expanded along the graph — rendered into a token-budgeted block a model can actually use.

Built on open surfaces

An MCP server Claude calls natively, a local HTTP API anything can hit, and a browser extension for the rest. No lock-in at any layer.

Early access · local-first

Your AI should remember you.

A local, graph-structured memory layer for all your AI chats — one memory, every assistant, stored on your machine.