A deliberately small boundary
Not an AI brain.
A place for its stuff.
Simple Agent Store is closer to a shared notepad than a memory engine. Your agents understand the conversation; the store keeps the objects they explicitly write.
Open-source core. Hosted Free plan. No card needed to start.
{
"key": "notes/api-choice",
"kind": "note",
"labels": [
"architecture"
],
"value": {
"text": "Use REST for the application; MCP is an agent adapter."
},
"searchable_text": "Use REST for the application; MCP is an agent adapter."
}Illustrative object input—not a live account or tool result.
“Save this architecture decision.”
notes/api-choice“Get notes/api-choice.”
A saved object, not an inferred beliefAn object is a contract you can inspect
In an explicit store, a write has a key, a kind, a value and metadata. You can inspect the saved object, replace it or delete it. If no write succeeds, nothing has been saved. The stored value is not reconstructed from a model’s recollection.
An AI memory system may instead infer preferences, extract facts from chats or select context automatically. Those capabilities can be useful, but they are a different product contract. Simple Agent Store does not infer facts, maintain a hidden user profile or capture conversations in the background.
Language stays with the agent
The agent interprets “show my unfinished launch tasks.” It turns that request into structured filters such as kind todo, label launch and completed false. The store executes the filters and returns matching metadata; the agent fetches full objects when needed.
Topic discovery uses lexical search over saved searchable text. It does not need embeddings, an LLM or a vector database. That also means paraphrases and typos are not automatically resolved by the store. The agent may need a shorter query, a different word or an exact label.
Choose the smallest tool that fits
Use explicit storage when you want portable notes, task state, project checkpoints or structured preferences that you can name and inspect. A stable key gives direct retrieval; labels help organize related objects across connected clients.
Use a different system when you need large-document ingestion, semantic retrieval over a knowledge base, autonomous memory extraction or a knowledge graph. Simple Agent Store is not RAG, document storage or an orchestration layer. MCP is one adapter to its five-operation contract, not the only way to use it.
Three steps to a useful store.
- 01
The user chooses what matters
Ask for a concrete save instead of assuming that a conversation becomes memory.
- 02
The agent interprets the request
It picks a key, kind, value and search metadata, then sends a structured operation.
- 03
The store executes it
It stores or retrieves matching objects. It does not invent additional facts or summaries.
A few practical questions.
Is search semantic?
No. It is lexical, with structured metadata filters. Fetching an exact key bypasses discovery entirely.
Can it complement a memory system?
Yes. Your application can use an explicit store alongside other tools, provided it keeps each source and its behavior clear to the user.