Coredence

Guide

AI memory tools compared: what each keeps, and where

Every assistant now promises to remember you. They mean different things by it — what is kept, who writes it, where it lives, and whether you can ever see why it believes something about you.

Ahmet Karaca ·

Four questions that separate them

  • What is kept: facts the model picked out of a conversation, notes someone wrote, or a record of what happened.
  • Who writes it: you, the assistant while it works, or a model summarising afterwards.
  • Where it lives: on your machine, or in the vendor's cloud behind an account.
  • Whether a memory can show its source: if it cannot, a wrong one is hard to spot and harder to argue with.

The notes below answer those for each tool, as of October 2026.

Memory built into the assistant

  • ChatGPT keeps saved memories (what you ask it to remember) and draws on past chats, in OpenAI's cloud. A newer system rolling out from June 2026 shows "memory sources" behind a reply, with OpenAI's own caveat that they may not show every factor. It works inside ChatGPT only.
  • Claude (claude.ai) builds memory automatically from your chats, with a separate memory and summary for each project, in Anthropic's cloud. You can view and edit it under Settings › Memory; incognito chats are left out.
  • GitHub Copilot Memory keeps facts about a repository and preferences about you in GitHub's cloud. Repository facts carry citations to the code that supports them and are checked before use. Anything unused for 28 days is deleted.
  • Cursor removed its Memories feature in version 2.1; its persistent context is now Rules, which you write.

All four are convenient and none of them leaves the vendor: what ChatGPT has learned about you does not reach Claude, and the reverse.

Tools you add

  • claude-mem captures Claude Code sessions automatically through lifecycle hooks and has a model compress them into summaries, stored in a local SQLite database with a local vector index. Open source (Apache-2.0); an optional paid cloud sync exists.
  • Mem0 extracts memories from conversations with a model call, as a library, a self-hosted server or a hosted platform. Open source (Apache-2.0). Its self-hosted OpenMemory was retired in 2026; the code is archived.
  • Basic Memory keeps notes as plain Markdown files on your disk, written by the assistant through MCP and editable by you in any editor. Nothing is captured in the background. Open source (AGPL-3.0), with a paid cloud.
  • Pieces captures continuously on your device — copied code, screens you view, audio — and offers it to Cursor, VS Code, Claude and ChatGPT through MCP. Paid; no open-source licence is published.
  • Supermemory extracts facts from conversations and syncs sources like Google Drive, Gmail and Notion, through a hosted API or a self-hosted binary. MIT licensed.
  • Papez keeps what you tell your assistants in one graph that every connected MCP app reads and adds to, scoring each memory by relevance and by how often it comes back, and forgetting what fades. The engine is open source (AGPL-3.0) and runs locally; a hosted version adds sync.
  • NoteMCP is a notes app your assistants can search and write to through MCP: links, screenshots and voice memos saved from your phone, kept as plain Markdown on its servers and exportable. Proprietary, with a readable MCP server.

What none of them quite does

Most of these remember the work: the project, the facts, what was said. Few try to describe the person doing it — how you decide, what you will not trade away, how you want to be spoken to — and fewer still keep each claim tied to the moments it came from, so that you can disagree with it about something specific.

That matters because a memory you cannot inspect is followed anyway. A preference recorded from one bad afternoon shapes every session after it.

Where Coredence differs

Coredence reads what you have already written — notes, a journal, your chats with ChatGPT and Claude, your Claude Code and Codex sessions, commits if you want them — and consolidates them into memories about how you work, not about a project. Every memory keeps the traces it grew from; a memory seen once waits as a draft; two memories in conflict become a question for you; and you can correct a memory or say it is not you.

The mind is one SQLite file on your machine, with no account. The same mind reaches Claude Code, Cursor, Codex, Gemini and any MCP client through a read-only server. The only call that leaves your machine is to the model you choose for consolidation.

Assistants only read the mind; none of them writes to it. You do: through what you write, and a line you drop from the menu bar. Claude, ChatGPT and Gemini chats take it through their own memory import, in one paste.

You can try the real app on a made-up person's mind, in your browser, without installing anything.

Try it in your browser