How to Give ChatGPT, Claude, and Other AI Tools One Shared Memory

Why every assistant forgets differently

Most people now use several AI assistants: one for writing, one for coding, one for search-heavy research. Each has its own memory feature, or none. Each decides on its own what to remember, stores it in its own format, and shows you only part of it.

The result is repetition. You explain your role, your project, and your preferences to each tool, then again when a memory silently drops something. What one assistant learned about your codebase is invisible to the one you use for planning. And none of it moves with you when you switch tools.

The shared-memory pattern

The fix is to move memory out of the assistants and into a store you control. The store holds your context. Each assistant connects to it and retrieves what it needs, when it needs it.

MCP makes this practical. A memory server exposes tools such as search memory and save note. Any assistant that supports MCP, or a connector bridging to it, can use the same server. Desktop clients and coding agents generally support MCP directly; web assistants vary and change often, so check each tool's current connector options.

Because the store is searched rather than loaded wholesale, it can grow large without every conversation carrying it. The assistant pulls the three notes relevant to the current question, not your entire history.

Setting it up

1. Choose the store. A folder of markdown notes, such as an Obsidian vault, is a good base: readable, portable, and yours. A dedicated memory service is an alternative if you prefer not to manage files.
2. Index it. Put the notes behind a retrieval layer so they can be searched by meaning, not just by filename.
3. Connect each assistant. Add the memory server to every client that supports MCP. Restart each, and confirm the memory tools appear.
4. Tell each assistant to use it. A line in each tool's custom instructions, such as an instruction to search memory before answering questions about ongoing projects, makes retrieval a habit rather than an accident.

What to put in shared memory

The quality of shared memory depends on what goes in, and automatic capture of every conversation produces noise: half-formed ideas, abandoned plans, and contradictions.

Good entries. Stable facts about you and your work. Preferences for tone, format, and tools. Project briefs with goals, constraints, and current status. Decisions and the reasons behind them. Reference material you look up repeatedly.

Poor entries. Raw chat transcripts. Temporary task details. Anything you would not want every assistant to read.

Review the store occasionally. Update project status, delete what is stale, and merge duplicates. Stale memory is worse than no memory, because assistants treat it as current.

Where shared memory goes wrong

• Assistants do not search it. Connecting the server is not enough. Without an instruction to search memory for relevant questions, many assistants answer from their own context and never call the tool.
• Conflicting memories. If an assistant's built-in memory says one thing and the shared store another, answers become inconsistent. Turn off built-in memory in tools where the shared store should be the only source, or keep built-in memory to trivial preferences.
• Everything is one note. A single giant document retrieves poorly. Split by topic and project so search returns the relevant piece.
• Nobody writes to it. Memory that is only read goes stale. Make saving a deliberate step at the end of meaningful work, such as a short note of what was decided and why.

Privacy and where the memory lives

A shared memory is also a single place that describes you, your work, and your clients. Decide where it lives and who can read it before filling it.

A local vault keeps notes on your machine, but retrieved passages still go to whichever model the assistant uses. A hosted memory service is convenient, but you are trusting its storage and access controls. The strongest setup keeps notes as private, encrypted chunks that only your connected agents can retrieve, and sends models only the passages a question needs, rather than whole files.