NeuroAIgent publishes this page and makes Central Brain. The comparison below is based on each vendor's official documentation and pricing pages as reviewed on October 28, 2026. It is not a benchmark, and no tool was ranked by retrieval accuracy.
Coding memory problems fall into three groups: rules and conventions the assistant should always follow, decisions and facts learned during the work, and source material (specs, PDFs, exports) the assistant should search. Built-in instruction files handle the first group well. The tools below differ mostly in how they handle the second and third.
Which memory tool is best when?
| Tool | What it is | Best when |
|---|---|---|
| CLAUDE.md and Claude Code auto memory | Instruction files loaded every session, plus notes Claude Code writes per repository | You work mainly in Claude Code and want zero setup |
| GitHub Copilot instructions and Copilot Memory | Repository instruction files; Copilot Memory (public preview) keeps repository facts and user preferences | Your team works in GitHub and Copilot agents, review and CLI |
| ChatGPT memory | Saved memories and a memory summary managed in ChatGPT settings | Your context is mostly personal preferences inside ChatGPT |
| Mem0 and OpenMemory MCP | Memory engine with a managed platform, Apache-2.0 open-source engine and an OpenMemory MCP guide | You are adding user or agent memory to an application |
| Zep and Graphiti | Managed context service; Graphiti is its Apache-2.0 temporal context-graph framework | You need facts that change over time, tracked per user, at production scale |
| Letta | Platform for stateful agents with persistent memory; Apache-2.0, self-hostable | You want agents that keep and edit their own memory |
| Cognee | Apache-2.0 memory engine that turns documents, code and conversations into a knowledge graph; includes an MCP server | You want graph-structured memory you can run yourself |
| Basic Memory | Plain Markdown knowledge graph over MCP; free open-source local version, paid Cloud | You want a readable personal knowledge base Claude can edit |
| Supermemory | Hosted memory and context API with an MCP endpoint and shared spaces | You want hosted memory shared across apps and teammates |
| Pieces | Desktop app that captures work activity into long-term memory saved on-device, with an MCP server | You want automatic capture of what you worked on |
| Central Brain | Local desktop app: indexes selected project files and keeps readable project notes, served over MCP | Your project evidence lives in local documents and you switch between AI clients |
Start with the built-in memory you already have
Claude Code loads CLAUDE.md files at the start of every session from the project (./CLAUDE.md or ./.claude/CLAUDE.md), your user folder (~/.claude/CLAUDE.md) and a private CLAUDE.local.md. Its auto memory writes notes per repository under ~/.claude/projects/<project>/memory/, with a MEMORY.md index whose first 200 lines or 25KB load each session. Anthropic recommends keeping each CLAUDE.md under about 200 lines.
GitHub documents repository custom instructions and Copilot Memory. Copilot Memory stores repository-scoped facts and user-level preferences, deletes entries unused for 28 days, and is used by the Copilot cloud agent, code review and CLI. If one assistant and one repository cover your work, these built-ins are usually enough and cost nothing extra.
Memory frameworks for agents you build
Mem0, Zep, Letta and Cognee are developer infrastructure: you call them from code to give an application or agent memory. Mem0 offers a managed platform and an Apache-2.0 engine; Zep is a managed service built on Graphiti, which tracks how facts change over time; Letta builds stateful agents that manage their own memory; Cognee builds a knowledge graph from documents, code and conversations. Pick among them by data model (vector, temporal graph, agent-managed blocks), hosting and the languages your stack uses.
These are a better choice than a desktop app when end users of your product need memory, when you need per-user isolation at scale, or when memory must run inside your own cloud.
Personal and project memory over MCP
Basic Memory, Supermemory, Pieces and Central Brain connect memory to assistants you already use. Basic Memory keeps notes as Markdown with typed relations; Supermemory is hosted with shared spaces; Pieces captures work activity automatically. Central Brain prepares PDF, DOCX, TXT, Markdown, CSV and XLSX files locally, offers semantic and exact search with source path, page and chunk where available, and keeps five note types plus a project brief as readable local files. Claude Desktop, Claude Code, GPT Desktop, GitHub Copilot in VS Code, Hermes and OpenClaw, connected over standard MCP with a configuration the app generates for each client.
A fair test for any of these: use one real project, ask a question with a known answer, change a decision, start a new session and check that the assistant retrieves the corrected decision and its source.
Common questions
Do I need a memory tool if I use Claude Code?
Often not. CLAUDE.md files and auto memory cover project rules and learned preferences in one repository. An external tool helps when memory must be shared across clients, machines or people, or when the assistant must search large document sets.
Which memory tools are open source?
Per their repositories, the Mem0 engine, Graphiti, Letta and Cognee are Apache-2.0, and the local version of Basic Memory is open source. Central Brain is a commercial desktop app.
Is this list independent?
No. NeuroAIgent publishes this page and makes Central Brain. Competitor descriptions come from their official documentation, linked below.
Next steps
Sources and review notes
- Claude Code: how Claude remembers your project
- GitHub: Copilot Memory
- GitHub: repository custom instructions
- Mem0 repository
- Mem0: OpenMemory MCP guide
- Graphiti repository (Zep)
- Letta repository
- Cognee repository
- Basic Memory
- Supermemory MCP
- Pieces MCP
- Central Brain product
- MITechNews: NeuroAIgent builds a Central Brain (September 22, 2026) — Press coverage, not independent testing.