Claude has a memory tool now. It is files on a disk. It still is not your knowledge base.

Anthropic gave Claude a memory tool. The interesting part is not that Claude can remember. It is how.

The agent stores what it learns as plain files in a directory. It reads them back when it needs them, writes new ones as it goes, and keeps the rest out of its active context until the moment it is useful. No vector database. No special memory format. Files in a folder, named and organized, that the model reads and writes like any other file. The tool is client-side: Claude requests the reads and writes, and your application stores the files wherever you choose.

We have been saying that is the right shape for two years. So first, credit where it is due. Then the honest part, because the memory tool and a knowledge base are not the same thing, and it is worth being clear about which job each one does.

Here is the short version. Claude's memory tool holds what the agent decided to remember while doing a task. A knowledge base holds what you decided to write down. One is the assistant's scratchpad. The other is your source material.

What the memory tool is good at

Give it credit first. The design is sound.

• It is just-in-time. Instead of stuffing everything into the context window up front, the agent records what it learns and reads it back on demand. Long-running sessions stay focused instead of drowning in their own history.
• It is plain files. Memories are text documents in a directory. Claude reads and writes them with simple file operations, the same way it would handle any other file.
• It is yours to store. Because the tool is client-side, the files live on infrastructure you control, not inside a black box. You decide where they sit and you can read them.

If you are building an agent that runs for hours and needs to accumulate state as it works, this is a genuinely good primitive. It solves a real problem: an agent that forgets what it figured out three steps ago is useless.

So we are not going to pretend it is not useful. It is. It also proves a point we care about. When the team that built the model needed somewhere durable to keep memory, they reached for files, not a black box. Plain, readable, ownable files. That is the bet a knowledge base makes too.

What it is not

The memory tool remembers things the agent learned. That is not the same as holding the things you wrote down.

Three gaps, and none of them are flaws. They are just what agent memory is.

The agent writes it, not you

The memory tool fills itself. The agent decides what is worth keeping while it works, and it writes the files. That is exactly what you want for an agent grinding through a task. It is the wrong owner for the knowledge you care about.

A runbook, a customer interview, a design decision, a style guide: those are things you author on purpose, in your words, because you decided they matter. You do not want a model's running impression of them. You want the exact text, the one you wrote, that you can open, edit, and trust next month.

A knowledge base is notes, folders, and tags in Markdown that you write and own. The memory tool is a workspace the agent keeps for itself.

It lives inside one agent

The memory tool's files belong to the agent you built with it. They are scoped to that one application. They do not follow you to a chat with Claude on the web, and they certainly do not follow you to ChatGPT, Cursor, or Copilot.

A knowledge base does the opposite. You write notes once, and any AI you connect through MCP reads the same ones. Ask Claude to draft from your notes. Switch to ChatGPT to send it, and ChatGPT reads the same notes. One store, every client. The assistant changes. The source stays put.

That portability is the whole point. Your knowledge should outlive your choice of model.

It is task state, not knowledge

This is the core of it. The memory tool makes one agent better at finishing one job. It is not where your team's architecture notes, your literature review, your interview transcripts, or the decisions you will need to quote next quarter belong.

That is the difference between a scratchpad and a system. The memory tool is a scratchpad inside one agent's run. The notes worth keeping belong in a system that any AI can read and that survives when a model, a workspace, or a roadmap changes.

Use both

This is not one-replaces-the-other. If you build agents, use the memory tool. Let the agent keep its working notes, accumulate state across a long task, and stay focused. That is what it is for.

Keep the knowledge worth writing down somewhere your AI can actually read it. Not a summary the agent inferred. The real documents, authored by you, in a place Claude and ChatGPT can both reach.

The agent's memory is the agent's. Your knowledge base is yours.

Who should pick what

The memory tool is enough if: you are building a Claude agent with it, you need it to carry working state across a long-running task, and the thing being remembered is the agent's own progress, not source material you will reuse elsewhere.

You want a knowledge base if: you use Claude or ChatGPT (or both, or plan to), you want notes you can read, edit, search, and export like normal documents, you want the same knowledge to serve every AI you use, and you care that it survives if a model or a workspace changes.

Give your AI a memory you own

Anthropic proved the shape: memory is files you can read. A knowledge base takes that the rest of the way. You author the files. You own them. Every AI you connect reads them.

Hjarni is a knowledge base with a built-in MCP server. You write notes in Markdown. Claude and ChatGPT read them, follow your instructions, and remember what you told them across every conversation. That persistent, searchable store is what long-term memory for AI actually looks like. Write once. You both remember.

Wiring it up takes about five minutes. Give Claude long-term memory walks the Claude path and the Claude setup guide has the connector steps, connect ChatGPT covers the other side, and the best ways to give your AI memory lays out every option, honestly. If you came here from the ChatGPT side, ChatGPT memory got an upgrade and it still is not a knowledge base is the sibling to this one.

Claude can keep its own working notes now. That is a good upgrade for agents. It is still not the place to keep what you write down.