Memory

With memory on, the model records facts the person shares, such as who they are, where they live and what they prefer, and reads them back on later turns. What it learned stays with it even after older messages no longer fit in the model’s context.

Memory belongs to one conversation. Nothing recorded in one conversation is seen in another.

The Node and Python examples use a client set up for this API.

Turn memory on

Send helpingio:memory: true on the first response of a conversation. Later turns need nothing extra: the conversation keeps memory on.

curl

curl https://api.helping.io/v1/responses \
  -H "Authorization: Bearer $HELPINGIO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "anthropic/claude-sonnet-5", "input": "I just moved to Lisbon.", "helpingio:memory": true}'

Node

const response = await client.responses.create({
  model: "anthropic/claude-sonnet-5",
  input: "I just moved to Lisbon.",
  "helpingio:memory": true,
});

Python

response = client.responses.create(
    model="anthropic/claude-sonnet-5",
    input="I just moved to Lisbon.",
    extra_body={"helpingio:memory": True},
)

A conversation you created with POST /v1/conversations gets memory the same way: send the field on its first response.

What you can rely on

• Memory is chosen once. A conversation keeps what its first response chose. Sending the same value on a later turn changes nothing; sending a different one is a 400 naming helpingio:memory.
• It is not accepted on POST /v1/conversations. Sending helpingio:memory there, with any value, is a 400 naming the field. Set it on the conversation’s first response.
• Memory needs a model that calls tools. A turn with memory on, sent to a model that cannot call tools, is a 400 naming model, and nothing runs.
• Your tool-call cap is yours. Recording or reading a fact never counts toward max_tool_calls, and the cap never cuts it.
• The model’s own memory work is not in the response. You see the reply, not the facts it recorded.
• Recording a fact is a model turn. The model records a fact with a tool call, so one more model turn follows. It is billed like any other and is included in the response’s usage.
• Deleting the conversation deletes what it remembered. A deleted conversation cannot be continued, so its facts are never read again, and they are erased with the rest of its content 30 days after the delete.