Agent Memory API Now in Beta on Telnyx

1, Oct 2026

Telnyx Agent Memory (Beta) gives AI agents long-term memory through one API. Each user, caller, or agent gets a profile. Your app writes sessions and facts in, and the agent recalls the ones that matter when the next conversation starts.

What's new

  • A profile per user, caller, or agent: Key memory to whatever id you already use, such as a phone number or a user ID. One profile never sees another profile's memories.
  • Ingest a session: Send a finished conversation in whatever shape your framework produces, from a voice call, a chat, or your own app, and Telnyx extracts the facts. Sending the same session_id again replaces that session, so a retry after a timeout is safe.
  • Remember one fact: Store a sentence exactly as written, such as a CRM field or a preference the user set, with no extraction.
  • Recall: Ask a question and get ranked facts back. No LLM sits in the recall path.
  • Summary card: Read one precomputed summary of a profile at the start of a session, with no query needed.
  • See and delete what is stored: List a profile's memories and sources, read a source as it was sent, forget one source, or forget a whole profile.

Why it matters

  • An agent without memory asks the same questions in every conversation. Agent Memory keeps what a customer told you on the phone available when they message you weeks later, on any channel.
  • Building memory yourself means an extraction job, a store, and a delete path that reaches derived facts. Agent Memory is one API on api.telnyx.com, called with the same API key as the rest of your Telnyx account, so there is no separate vendor to integrate.
  • Every memory traces back to the source it came from, so you can show a user what the agent knows and remove it.
  • During Beta, writes are asynchronous: a write returns 202 with an operation to poll, and a memory is recallable once that operation completes, usually within seconds. One namespace, default, is available today.

Example use cases

  • A booking agent that recalls a customer's seat and meal preferences from a phone call when the same customer messages in chat weeks later.
  • A voice agent that greets a repeat caller with what it learned on the last call instead of starting from zero.
  • A sales assistant that follows a prospect across calls and chats without re-asking the same qualifying questions.
  • A personal assistant that keeps preferences current as they change.

Getting started

  1. Use your existing Telnyx API key. There is nothing to provision: the default namespace and each profile are created on first use.
  2. Write a fact to a profile:
curl -X POST "https://api.telnyx.com/v2/ai/memory/namespaces/default/profiles/user_4417/remember" \
  -H "Authorization: Bearer $TELNYX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"text": "Prefers an aisle seat and a vegetarian meal"}'
  1. Poll the returned operation_id at /namespaces/default/operations/{operation_id} until its status is completed.
  2. Recall what the agent needs:
curl -X POST "https://api.telnyx.com/v2/ai/memory/namespaces/default/profiles/user_4417/recall" \
  -H "Authorization: Bearer $TELNYX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query": "What seat and meal should I book?", "top_k": 3}'
  1. When a user asks to be forgotten, delete the profile with DELETE /namespaces/default/profiles/{id}, or delete a single source to remove just the memories it produced.

Learn more in the Agent Memory docs, including writing memories, recalling memories, and deleting memories.