Where it sits in the loop
- 1. Query arrives
- A user prompt, a tool call, or a scheduled trigger.
- 2. Recall and search
- Antfly embeds the query, runs hybrid search, and reranks.
- 3. Ground and reason
- Context and memory go to the model; tool calls come back.
- 4. Store and return
- Antfly stores the exchange and updates the entities it mentions; the answer streams back.
Steps 2 and 4 are one engine and one table, so the memory an agent writes is what the next query searches.
What one call does
One request embeds the question, runs keyword and vector search together, reranks the top candidates in a single batched call, and prunes the weak results. Put three help articles in a table and ask whether you can get your money back after canceling: the cancellation and refund articles come back scored, and the billing-cycles article never reaches the model. The support answer agent guide runs exactly this, then writes the answer with a local model.
Shared access
Several agents on one team read the same tables, and not all of them should see the same rows or be allowed to write.
- Roles
- On a self-hosted node, enable_auth turns on authentication and RBAC. The CLI's auth commands manage users, roles, permissions, and API keys.
- Grants
- In Antfly Cloud a key or a group is granted by table, operation, and role level, so one agent can write to its memory table and only read the shared one.
- Row filters
- A row filter on the key or group limits retrieval to the rows it may see.
- Tool visibility
- Over MCP, a tool the key cannot use is not listed to the agent.
Memory
MemoryAF, a showcase project built on Antfly, is a memory service: an MCP server, a REST API, and a dashboard over hybrid search with entities and relationships. Any MCP client can store to it and recall from it, so agents share one memory instead of each starting from zero.
Local to distributed
In standalone mode and in Lite, embedding and reranking run in the same process as the query; in a cluster the inference pool sits beside the data. With Antfly Inference as the embedder, a query never leaves the process or the cluster to embed or rerank. Antfly Lite opens the database file directly from a Go or Zig application, with no server between the agent and its data. The same tables, indexes, and queries move from that file to a standalone node to a cluster without changing the agent.
How to talk to it
Start with the REST API or an SDK. Add MCP when an agent should discover tools instead of being handed them.