Lou Montulli had the idea of using cookies in web communications in June 1994, while Netscape was building an e-commerce application for MCI. HTTP kept no state between requests, and a cookie gave a shopping cart somewhere to live.
A model call is stateless in that way too, so what an agent should know from last week has to be put into the request. A cookie holds a value the server wrote on purpose, though, and agent memory is distilled from conversation, where a later fact can contradict an earlier one.
I ran a team of 13 writers at Mem0, which sells a memory layer, so weigh this topic with that in mind.
Memory is a pipeline the application owns from the write to the read, and each stage has its own failure.
What memory is
- AI Memory Management for LLMs. Covers the memory layers and who decides what gets written, with runnable examples of eviction and scope.
- Episodic, Semantic, and Working Memory in AI Agents: A Practical Map. Maps the three memory types an agent uses to what each one stores.
- Context Windows vs Memory: Why They Are Not the Same Thing. Treats the context window as a per-request budget and memory as state retrieved into a later request.
- RAG vs Memory: What AI Developers Need to Know. Draws the line between retrieval for one request and state kept across requests, and shows how to run both.
Short-term memory and compression
- Short-Term Memory for AI Agents: A Practical Guide. Covers what compaction keeps and loses in the context window, with a tested example of a dropped instruction.
- Contextual Compression for Agent Memory: What Stays and What Goes. Compares three compression strategies for deciding what stays when context is finite.
Persistence and audit
- Memory Serialization: How Agents Persist State Across Sessions. Explains why an agent forgets on restart and the serialization patterns that carry state across sessions.
- Memory Versioning and Audit Trails for Regulated AI Agents. Builds append-only memory versioning so an agent's reasoning can be traced in an audit.
- I went looking for the citation field in a proposed agent memory format. Argues that a memory record should carry how the agent knew, ahead of any shared format.
Several agents, and voice
- Shared Memory vs Isolated Memory in Multi-Agent Workflows. Chooses between shared and isolated memory for multi-agent systems by coordination and debugging cost.
- Memory for Voice AI Agents: What Text Chatbots Cannot Do. Covers memory across a real-time streaming pipeline, where a voice agent has constraints a text chatbot lacks.
The field and its tools
- State of AI Agent Memory in 2026. Surveys agent memory architectures in 2026 and ties the choice to the agent's task and risk.
- How Memory Works in HyperAgents. Walks through how HyperAgents layers memory and keeps session continuity.
- How Memory Works in DeerFlow. Walks through DeerFlow's memory, from session state files to sub-agent isolation.
Does an agent need memory if it has RAG and a large context window?
Yes, when state has to survive across requests. The page on context window vs memory treats context as a per-request budget, and RAG vs memory shows that RAG selects evidence for one request.
Related topics
A memory read path is a retrieval problem, and the RAG pipeline topic covers the search stack it runs on.
Context window limits are the reason memory exists, and the LLM inference topic covers what a long context costs.
Claude Code's memory files are one product's version of this design, covered in the coding agents topic.