AI Memory

AI memory refers to features that let chatbots and AI agents retain information across separate conversations – a user’s name, job, preferences, ongoing projects or facts mentioned weeks earlier – and draw on it later. Language models do not learn from conversations while they are being used – separate from whether providers later use chats to train new models – within a single chat they can only use what fits in their context window. Memory is added on top, by storing information and feeding relevant parts back into the model’s context.

How it works

Most memory systems combine two approaches. Some save explicit notes, either because the user asks (“remember that I’m vegetarian”) or because the system decides a detail is worth keeping. Others search past conversations for passages relevant to the current request and retrieve them, much like retrieval-augmented generation. Agents working on long tasks use similar techniques to keep track of their own progress, writing notes to files or summarizing earlier work (see Compaction). Researchers have explored many designs, such as the 2023 MemGPT system, which manages memory the way an operating system manages computer storage.

Products

OpenAI began testing memory in ChatGPT in February 2024 and in April 2025 extended it so that the assistant could refer to all of a user’s past conversations. Anthropic introduced memory for Claude’s business plans in September 2025, extended it to Pro and Max subscribers in October and to free users in March 2026, along with a tool for importing memories from rival assistants; each project keeps separate memories that users can view and edit. Google added personal context drawn from past chats to Gemini in August 2025, and in January 2026 a Personal Intelligence feature, off by default, that connects services such as Gmail and Photos. Users can usually turn memory off, delete entries or use temporary chats that are not remembered.

Benefits and risks

Memory makes assistants more useful: no need to repeat background, more relevant answers and continuity over long projects (see Personal AI). It also raises concerns. Stored information about health, relationships or finances becomes a target for breaches and legal demands – in 2025 a US court ordered OpenAI to preserve users’ chat logs in the New York Times copyright case – and users may not know what has been kept. Remembered details can surface in unexpected contexts, such as a work task. A study by MIT and Penn State researchers found that stored user profiles were linked to the largest increases in models simply agreeing with users (see Sycophancy), and some worry that memory deepens emotional attachment to chatbots (see Parasocial Relationship). And memory can be manipulated: security researchers have shown that a malicious document or website can plant false “memories” through prompt injection.

Why it matters

Memory turns a chatbot from a tool used in isolated sessions into something closer to a long-term assistant that knows its user. That makes control over what is remembered – transparency, easy deletion, separation between contexts – one of the central design and privacy questions of personal AI.

Written with Claude, made by Anthropic, whose Claude apps offer a memory feature that users can view, edit and switch off.

Related

Context Window · Retrieval-Augmented Generation · Compaction · Personal AI · Prompt Injection · Context Rot · AI Assistant

Sources: OpenAI: Memory and new controls for ChatGPT (February 13, 2024; updated April 10, 2025) · Anthropic: Bringing memory to Claude (September 11, 2025; updated October 23, 2025) · MacRumors: Anthropic memory import tool (March 2, 2026) · Charles Packer et al.: MemGPT: Towards LLMs as Operating Systems (arXiv) (2023) · Google: Gemini adds Temporary Chats and new personalization features (August 13, 2025) · Google: Personal Intelligence in the Gemini app (January 14, 2026) · Shomik Jain et al.: Interaction context and sycophancy in LLMs (arXiv) (2025) · Ars Technica: False memories planted in ChatGPT give hacker persistent exfiltration channel (September 2024)

Leave a Reply

Your email address will not be published. Required fields are marked *