Twilio Remembers Nothing Between Calls. Its New Conversation Memory Changes Who Does the Remembering.

You shipped a voice agent on Twilio. During the call it is sharp. It picks up context mid-conversation, handles interruptions gracefully, pulls the right answer from the knowledge base. Then the caller hangs up, calls back the next day, and says: "Hi, it's me again. Did that refund go through?"

The agent has no idea who "me" is.

It feels like a bug. It is not a bug. It is the architecture doing exactly what it was designed to do. The honest answer to "does Twilio remember between calls" used to be a flat no, and the honest answer today is: it depends entirely on which Twilio you bought, and even then, the memory lives inside Twilio's world, not yours.

ConversationRelay: the transport remembers nothing

Start with what most builders actually use. Twilio ConversationRelay connects a phone call to your WebSocket server. It manages the hard real-time parts: speech-to-text (Google or Deepgram), text-to-speech (Google, Amazon, or ElevenLabs), barge-in when the caller interrupts, DTMF detection, partial transcripts. When a call connects, your server receives a setup message with the call SID, a session ID, and the caller and callee numbers. From that moment on, your server is the brain: you own the LLM, the prompts, the tools, and the history.

Community reference docs for ConversationRelay state the boundary plainly: no built-in memory or context, bring your own conversation history and context management; no LLM integration, pure transport layer. That is the deal. Twilio moves audio and text with impressively low latency. Remembering is your department.

The working pattern, visible across open-source ConversationRelay agent templates, is to key conversation history by the caller's phone number from the setup message, not by the call SID. The call SID dies with the call. The number survives it. So when the same caller phones back, your server loads the stored transcript or summary keyed to that number and injects it into the prompt. Some templates persist to a local SQLite file; others use Twilio Sync; the more polished ones add a memory retrieval mode that recalls context once per conversation to keep token costs down. All of them share one property: the schema is yours, the storage is yours, the read on every call is yours. The agent remembers exactly as much as your plumbing lets it remember.

For a single demo this is fine. For an operation running hundreds of calls a week, the DIY memory layer becomes its own product to maintain: what to store per caller, how long to keep it, how to summarize long histories, how to keep test calls out of production memory, how to make the same memory available to the parts of the business that are not phone calls at all.

The managed answer arrived in 2026

Twilio noticed the gap, and this year it shipped the managed answer. At its SIGNAL conference in May 2026, Twilio launched a "conversational layer" for the agentic era: Conversation Memory, Conversation Orchestrator, Conversation Intelligence, and Agent Connect, all generally available the same day after a private beta running since January. The headline capability does what the name says: it extracts and maintains customer history, preferences, behavior, and conversation state across every channel, so each conversation picks up where the last one left off instead of starting from zero.

Then on October 12, 2026, Twilio made the integration that matters for existing customers generally available: Segment Conversation Memory. If a business already unifies customer identity in Segment's CDP, those profiles now activate inside live conversations across Twilio Voice, SMS, WhatsApp, and Flex, without custom webhook pipelines. Twilio's own framing is precise: Segment is the long-term system of record for identity and consent; Conversation Memory is the sub-300ms working memory that puts that context in front of the agent, human or AI, while the customer is still on the line.

Twilio's urgency here is backed by its own research. Its 2026 Customer Insights Series found that 68% of Indian consumers have stopped using an AI agent mid-interaction because it lacked basic information about them or their previous interactions, and 75% say they have to repeat information they already shared. Brands, meanwhile, are overconfident: nearly 9 in 10 believe their AI agents effectively recognize returning customers. That gap between what brands believe and what callers experience is exactly what Conversation Memory is built to close.

Where the automation operator still gets burned

Here is the boundary that matters for anyone running scheduled automations, not just phone trees. Twilio's memory is scoped to Twilio's world: its channels, its customer engagement stack. It remembers the caller across voice, SMS, and WhatsApp. It does not remember what your n8n workflow decided about that customer at 2 AM. It does not remember the summary your scheduled research agent wrote yesterday, or the follow-up your Make scenario queued, or the context your Claude Code session gathered while fixing the billing integration.

The automation operator's real stack is multi-tool by definition. The same customer calls in, then shows up in a workflow run, then gets touched by a scheduled agent. If memory only lives inside the phone system, every other agent in the stack still wakes up blind. You end up with the exact failure mode this pipeline exists to kill: one channel that remembers, surrounded by agents that forget.

That is the question to ask before adopting any vendor's memory layer: who operates the memory your agents share across everything you run? A memory layer that only works inside one vendor's transport is infrastructure for that vendor's agents. The operator needs memory that follows the operator.

This is the problem Vilix AI is built for. It is cloud-hosted, so there is no infrastructure to run and nothing to provision per caller. The same memory is available everywhere over MCP: the voice agent, the scheduled workflow, the coding assistant on your laptop, the phone app. It stores full conversation history, not just extracted traits and summaries, so the reasoning behind a decision survives, not only its outcome. It is free forever on the free plan, with a 7-day Pro trial that asks for no credit card, and your data is portable: export everything or delete it anytime. Learn more at https://vilix.ai/?utm_source=vilix-blog&utm_medium=article&utm_campaign=twilio-remembers-nothing-between-calls-conversation-memory.

The question has changed

A year ago, "does Twilio remember between calls" had a one-word answer. Today the answer is a map: ConversationRelay itself persists nothing and never will, because it is transport; the managed conversational layer persists customer context across Twilio channels as of 2026; and the operator's cross-tool memory is still the operator's job to wire.

So the question worth asking is not whether Twilio remembers. It is whether your memory answers to one channel vendor or to you. The agents that win the next two years will be the ones that wake up knowing the customer, no matter which tool is doing the waking.