What you save is what stays.
The store keeps explicit objects. It does not build a profile or invent facts about you.
[ ] Open-source storage. Your agents’ stuff.
Dead simple
agent storage.
Save something with one agent.
Find it with another. Keep it yours.
One store.
All the little things.
↳ Notes✓ Todos@ Contacts{ } Any JSON
Say it. Save it.
Pick it up anywhere.
One useful thought. Saved in Claude, ready in Codex.
You
Let’s start with five teams.
Just chatting
agentstore
Your next note
A place for the useful bits.
Nothing saved yetA thought worth keeping.
A shared notepad. With an API.
Different agents.
Same drawer.
AgentStore holds the things you ask your agents to save. Connect them to the same store and they can pick up where another left off.
agentstoreEmpty drawer
A place for the thing you want to keep.
Pick up with Codex
Save an example above to see the handoff.
Just enough storage.
No extra intelligence.
A name. Some tags. Your data.
Give each object a key and kind. Add labels to find it again, or fetch it directly by key.
Your agents bring the brains.
They understand your words. AgentStore handles the storing, searching, and retrieving.
Store
Worth keeping?Give it a home.
The useful part. Not the whole conversation.
A decision, a todo, a contact, a few project settings. Ask your agent to save it as an object you can find and use later.
- A key to call it by
- Give each object a stable address, like notes/launch-decision.
- A little structure
- Kinds and labels keep notes, tasks, and references easy to sort.
- Your actual content
- Keep text or JSON. Update it when things change; delete it when you are done.
An example saved object
Try saving an objectnotes/launch-decisionnote / launch / planningStart with five teams. Expand after feedback.
Finding it is the point.
A clear way
back to your data.
Know the key? Get the object. Need everything? List the collection. Looking for a detail? Search words and metadata.
- Your agent interprets the request.It chooses search terms and filters such as kind or labels.
- The store returns matching candidates.Compact results show which objects might be relevant.
- Get brings back the full object.Your agent reads the saved value, not a generated summary.
Search is lexical. For paraphrases or typos, the agent may need to try different words.
Try:
2 candidatesChoose one to get its value
Small surface. Useful building block.
You already know
how this works.
Put, get, list, search, delete.
Five operations, whatever agent you use.
Give it a home.
Save an object under a key. Replace it when it changes.
put({
key: "notes/launch",
kind: "note",
labels: ["launch"],
value: { text: "Invite five teams first." }
})Bring your own agents.
One store to connect to.
Connect Claude Code, Codex, or another compatible MCP client to your Cloud store—or self-host the open-source core.
Find your connection guide- Create your store.
Start with a free Cloud account and create an access key in Get started. Prefer local storage? Run the core yourself.
- Connect each agent.
Cloud uses remote Streamable HTTP with bearer access keys. Point every compatible client at the same account’s store.
- Ask for a save.
Optional routing instructions help agents recognize save and retrieve requests. Check the object in the data viewer; client approvals still apply.
Apache 2.0
Your data deserves
a home you own.
Run the open-source core yourself, or use Cloud to manage the infrastructure. The same simple idea: explicit objects, clear retrieval.
Explore the open-source coreA few good questions.
Read the docsIs this an AI memory system?
No. It stores objects your agents explicitly write. It does not learn from conversations, infer facts, or maintain a hidden profile.
Will it save every conversation?
No. The connected agent must call a storage tool. A routing skill can help it respond to “save this,” but it does not guarantee a tool call.
Can I store more than text?
Yes. Objects can contain JSON values and carry keys, kinds, labels, versions, timestamps, and optional expiry. Contact details and settings fit too.
Does it need embeddings or an LLM?
The store needs neither. It uses structured filters and lexical search. Your agent handles language and decides what to do with the results.
Can I use the hosted service today?
Yes. Start with 100 free objects, connect a compatible agent and inspect your data in the web viewer. Paid plans add room, not per-agent fees. You can also self-host the open-source core.