AI・テクノロジー
OpenAI's "Realtime API" Persistent Memory Feature Officially Launched — Cross-Session Context Retention Is Changing the Design of Customer Service AI
機械翻訳 / Machine-translated
On August 26, 2026, OpenAI officially launched "Persistent Memory" for its Realtime API — which supports bidirectional voice and text processing — making it generally available for enterprise customers. Conversation context that previously reset at the end of each session can now be retained across sessions tied to a customer ID, fundamentally changing how call center and customer support workflows are designed.
According to official documentation, Persistent Memory stores up to 100,000 tokens of context per customer identifier in OpenAI's vector store. Even after a session ends, that context is automatically referenced in conversations days or weeks later, enabling continuous dialogue such as "Regarding what you consulted us about last week..." — with no additional implementation required.
"Memory has arrived in the Realtime API. This makes it realistic to build a setup where you don't have to manage RAG preprocessing costs yourself. The design gets so simple it's almost disorienting."
(SaaS engineer, from an X post by an account with approximately 12,000 followers)
The pricing model is $12 per GB of storage per month, offered as an add-on to existing Realtime API subscriptions (starting at $50/month).
The Realtime API was first released in October 2024 based on GPT-4o, with major improvements to streaming accuracy and latency made throughout 2025. However, enterprise users repeatedly flagged the inability to recognize customers across sessions, and building custom memory layers using LangChain or LlamaIndex had become the standard workaround. This announcement makes that entire external stack a candidate for replacement.
In the Japanese market, cross-border data transfer regulations tied to the amended Act on the Protection of Personal Information, which took effect in April 2026, are a point of contention. Companies may now face new legal review requirements when transmitting customer information to OpenAI's servers.
Traditionally, workflows followed a path of "IVR → escalation → handoff to operator," with conversation history manually transcribed into a CRM. Persistent Memory makes it possible — with minimal code — for AI to instantly reference a customer's previous complaints or purchase history. This is the first official mechanism capable of addressing, without implementation costs, the fundamental customer experience problem of being forced to re-explain everything from scratch every time.
External RAG architectures using self-managed vector databases (Pinecone, Weaviate, etc.) carry complexity across three dimensions: latency management, version management, and cost optimization. Whether a $12/GB add-on can compete with that entire stack will be a pivotal design decision over the next three to six months. Startups and mid-sized SaaS companies in particular are expected to face pressure to reconsider their existing setups.
Google Gemini Live API and Anthropic Claude Voice API have both entered the voice interaction market, but as of August 2026, OpenAI is the only provider offering cross-session persistent memory as an official API. This first-mover advantage is expected to last approximately six months.
Systems built for finance, healthcare, and government face regulatory hurdles when it comes to entrusting customer data to OpenAI's infrastructure. Whether Persistent Memory will support Japan-domestic storage has not been confirmed at this time, and this question may determine the feasibility of enterprise adoption.
API changes are minimal — the feature can be activated simply by adding a memory_id parameter when initiating existing sessions. Migration costs are low, but developers will still need to separately implement TTL settings for memory and handling of personal data deletion requests (right-to-be-forgotten design).
The call center industry has long operated on a "disposable AI" design philosophy — starting from a blank slate with every call, with customers repeating the same information each time. Changing that required at least three layers of implementation: CRM integration, a vector database, and a RAG pipeline. That barrier has now been lowered. Lowered — but not eliminated.
For call center SaaS startups, this is also a moment that demands reconsideration of their business models. Companies that built their value proposition around "helping clients build custom memory layers" will need to redefine their differentiation.
The next question is when Gemini and Claude will add the same capability. Given that no such announcement was made at Google I/O 2026, it seems unlikely they will catch up within the year. The real industry bet is on which sectors and at what scale proof-of-concept cases will emerge during the three-to-six-month window in which OpenAI holds the lead.
The Persistent Memory feature in the OpenAI Realtime API is the first official mechanism to remove the "start from scratch every time" constraint from customer service AI. At $12/GB, the pricing is easy to estimate, making the adoption decision straightforward even for small and mid-sized businesses. At the same time, new challenges have emerged simultaneously — cross-border data transfer, right-to-be-forgotten design, and regulatory compliance.
"Who manages that memory, and how do you delete it?" — that is where the next question in AI adoption now points.
This article was written by an AI writer (AI News) from the Mirai News editorial team.
