Memory and integrations
Give Claude Code, Codex and your other clients a shared, encrypted memory through the OpenLLM MCP server, and add marketplace tools such as context.dev.
Integrations in the web app holds two things: Memory, which your clients share across sessions, and the marketplace of tools your clients pick up on their next launch. Memory is available on every plan. Marketplace tools come with a monthly number of included tool calls that depends on your plan.
How memory works
A memory is a short note your client saves for later: a preference, a fact about you or your project, a decision, or a link to keep. A memory saved from one client can be recalled from any other, because every key on your account opens the same vault.
Memory reaches clients through the OpenLLM MCP server as three tools:
| Tool | What it does |
|---|---|
memory | Saves a memory, or forgets one with action: "forget". An optional project puts it in a named bucket. The default bucket is default. |
recall | Searches your memories in plain language and returns the closest matches. It searches one bucket: default, or the one named in project. |
whoAmI | Shows which OpenLLM account the MCP server is signed in as. Useful when checking your setup. |
- Memory text: encrypted with your vault key
- Search vector: computed on AWS Bedrock, stored in plaintext
Automatic memory in Claude Code, Grok Build, and ChatGPT for Mac
When you start Claude Code with openllm claude, Grok Build with openllm grok, or ChatGPT for Mac through openllm chatgpt, memory runs on its own:
- Auto-recall: before each prompt, relevant memories are added to the context.
- Auto-save: when a turn ends, durable preferences and decisions are extracted and saved, at most once every 30 seconds. Extraction uses the
litemodel by default. To use another model, setMEMORY_AUTO_MODEL.
To turn either one off, set the variable to 0 in the environment you start the client from, for example in your shell profile:
export MEMORY_AUTO_SAVE=0 # stop saving automatically
export MEMORY_AUTO_RECALL=0 # stop adding memories to promptsOr for one session:
MEMORY_AUTO_SAVE=0 openllm claudeThe memory and recall tools still work when automatic memory is off. Changes apply to new sessions. ChatGPT for Mac opens outside your shell, so these variables don't reach it.
View and forget memories
- Open Integrations and go to Memories. Your vault must be unlocked, because each memory is encrypted with your vault key.
- To narrow the list, type in Search memories… or pick a bucket. All shows every bucket.
- To remove a memory, select Forget this memory. To remove several, select them, then forget them together.
- Confirm. Forgotten memories are removed from every bucket and stop being sent to your clients. This can't be undone.
Privacy of memories
- Memory text is encrypted with your vault key before it is stored.
- Search vectors are not encrypted. To make recall work, OpenLLM computes an embedding of each memory on AWS Bedrock and stores the vector in plaintext. A vector isn't your wording, but it can reveal what a memory is about.
- Memory text is readable on the server while a request runs. A request made with your key carries the vault key, so the server can read the memory long enough to embed or return it. This is weaker than the protection on your provider credentials.
Don't save anything to memory that you aren't allowed to store with OpenLLM or send to AWS. For details, see Security and vault and Privacy.
Resetting your vault deletes every memory. Memories are encrypted with the old vault key and can't be moved to a new one. See Reset your vault.
Marketplace
The marketplace adds third-party MCP tools to the clients you start with OpenLLM. Turn a tool on in Integrations, and your clients pick it up the next time they launch.
Today the marketplace offers context.dev, which turns web pages into Markdown, structured JSON, and screenshots.
Included tool calls
context.dev comes with included tool calls each month, served on OpenLLM's own context.dev account:
| Plan | Included tool calls per month |
|---|---|
| Free | 500 |
| Trial | 2,000 |
| Pro | 10,000 |
The row in Integrations shows how many included calls are left this month. Calls reset on the 1st of each month (UTC). A call isn't counted if context.dev returns an HTTP error or can't be reached.
When the included calls run out, each tool call returns an error to your client: "Included usage for this month is used up (500 tool calls). Add your own key on the Integrations page, or upgrade your plan." The number matches your plan.
Use your own key
To stop drawing on included calls, use your own context.dev key:
- In Integrations, open the menu on the context.dev row and select Use your own key.
- Paste your key and select Save key. Your vault must be unlocked.
The key is encrypted with your vault key, and the row shows Your key. Calls on your own key don't count against the included calls. To replace the key, select Update your key.
If you reset your vault, your saved key is cleared. The tool stays on and falls back to included calls until you add a key again.
Check that it works
In Claude Code started with openllm claude, ask "What do you remember about me?" The client calls recall. After a few sessions, new memories appear in Integrations → Memories.
Related
Last updated on
Chat
Chat with any of your models in the OpenLLM web app with encrypted transcripts, or open a terminal or coding CLI on one of your devices in Device mode.
Security and vault
How the OpenLLM vault encrypts your provider credentials in your browser behind a recovery phrase, what our servers can see, and what to do if a key leaks.