How does AI agent memory work?
The short answer AI-agent memory lets an agent retain, retrieve, update, and use information beyond what fits in the model's immediate input window [Source](https://arxiv.org/html/2611.09721). It splits into short-term memory, supplied directly in the current context, and long-term memory, stored outside the context and retrieved when needed [Source](https://arxiv.org/html/2611.09721). Long-term memory is further divided into episodic, semantic, procedural, and external or tool memory, each with different content and retrieval methods [Source](https://arxiv.org/html/2611.09721). A memory-management layer governs what gets written, consolidated, retrieved, corrected, expired, or deleted [Source](https://arxiv.org/html/2611.09721).
How it works The model's context window is not the same as persistent memory: context is available only when included in the current request, whereas long-term memory is stored separately and selectively retrieved [Source](https://arxiv.org/html/2611.09721). Working or short-term context holds the current prompt, recent dialogue, tool results, and instructions, passed into the model's context window via conversation buffers, sliding windows, or summaries [Source](https://arxiv.org/html/2611.09721). Episodic memory stores past events, conversations, tasks, and outcomes, retrieved by semantic similarity, recency, metadata, or time, typically using vector databases, document stores, or temporal graphs [Source](https://arxiv.org/html/2611.09721). Semantic memory holds facts, preferences, and stable knowledge, extracted and updated by the system and injected into future prompts via fact stores, knowledge graphs, or key-value databases [Source](https://arxiv.org/html/2611.09721). Procedural memory stores reusable instructions and learned action patterns, retrieved or updated after task success or failure, using documents, code repositories, or policy stores [Source](https://arxiv.org/html/2611.09721). External or tool memory covers files, email, calendars, CRM records, web pages, and databases, queried through APIs, SQL databases, cloud storage, or search indexes rather than relying on the model's internal parameters [Source](https://arxiv.org/html/2611.09721).
A common retrieval pipeline: the agent receives a conversation, observation, or tool result; a memory-writing component decides whether it is worth retaining; the system stores raw text, a summary, an embedding, structured fields, or a graph relation; a later query is embedded or analyzed; candidate memories are retrieved by similarity, keywords, metadata, recency, graph traversal, or a hybrid method; a reranker or policy filters candidates; selected memories are inserted into the prompt; and the agent may revise, merge, expire, or delete memories afterward Source. Vector stores generally hold an embedding plus original content and metadata, enabling semantic retrieval but not, by themselves, truth verification or temporal reasoning Source.
Who makes and uses it OpenAI provides a memory feature described in its own help documentation, covering what memory is and how it functions within its products [Source](https://help.openai.com/en/articles/8983136-what-is-memory). Pricing and maker details beyond OpenAI's own documentation are not stated in sources.
Anthropic's Claude includes chat search and memory features intended to let the assistant build on previous context, according to Anthropic's support documentation Source. Pricing details are not stated in sources.
Standards and research work on agent memory is also underway: an IETF draft addresses agent memory integrity Source, with a related draft also covering agent memory integrity Source.
Numbers and dates - As of 2026-11-22, OpenAI documents a memory feature in its help center explaining what memory is and how it works [Source](https://help.openai.com/en/articles/8983136-what-is-memory). - As of 2026-11-22, Anthropic documents Claude's chat search and memory features for building on previous context [Source](https://support.claude.com/en/articles/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context%23h_a11816bb20). - As of 2026-11-22, an IETF draft on agent memory integrity is listed under the BMWG working group [Source](https://datatracker.ietf.org/doc/draft-khandelwal-bmwg-agent-memory-integrity/01/).
What is still open Vector stores enable semantic retrieval but do not, by themselves, provide truth verification or temporal reasoning, a limitation noted directly in the memory-architecture research [Source](https://arxiv.org/html/2611.09721). Standards work on agent memory integrity is still at the draft stage within IETF working groups, indicating the area is not yet settled [Source](https://datatracker.ietf.org/doc/draft-khandelwal-bmwg-agent-memory-integrity/01/). Details of how memory-management policies decide what is worth retaining, beyond the general description of writing, consolidating, retrieving, correcting, expiring, and deleting memories, are not further specified in sources [Source](https://arxiv.org/html/2611.09721). ## Latest on Agentry
• MinIO Launches AIStor Memory for Enterprise AI Agents
• SciMARS Memory Lifts Multi-Agent Science Benchmark Scores
• Couchbase launches managed agent memory and MCP server
• Microsoft Agent Framework adds memory, checkpoints in September releas
• Azure Databricks agent memory pipelines enter beta
• Oracle Select AI Agent adds configurable memory and team APIs


