Agent Memory
Agent Memory is a key technique for addressing the stateless nature of large language models. It uses external storage mechanisms to persist an agent's working context, decision history, and learned patterns, and supports shared read/write access and conflict resolution across multiple agents.
Core Facts
Timeline (last 90 days)
Anthropic calls the compression process in Agent Memory that turns raw conversation streams into usable context "dreaming."
Current engineering practice generally adopts a "frozen parameters + external memory read/write" architecture to manage Agent memory, rather than updating weights at inference time.
For developers, practicable approaches include: writing key facts into structured notes rather than relying on model memory; setting clear read and write rules for the agent to limit what it can modify; and actively verifying whether memory has been loaded correctly when switching sessions.
Memory loading is automatic, and users do not need to manually remind the AI what was discussed last time.
The system can set rules for the agent, requiring it to fill in missing necessary notes or maintain a work record in a prescribed format, which is called "automatic read control" and "note rules."
Automatic loading means the model must make trade-offs within limited context, and the quality of the retrieval strategy directly determines the accuracy of memory.
A memory system should not be merely a passive storage-and-retrieval pipeline, but can actively construct associative structures among memories and endow memories with flexible expressive capability.
The Agent's memory mechanism solves the stateless problem by persisting key information in external files; during operation the Agent writes valuable feedback, preferences, and lessons into the memory file, and reads it first on the next startup.
Well-designed memory requires filtering rules, retaining only feedback with universal value for future tasks and avoiding sedimenting one-off changes into permanent rules, otherwise the memory will bloat and produce conflicts.
The Agent Memory information officially disclosed by Databricks is limited, and details such as the storage granularity of memory, the recall mechanism, how it works with vector retrieval components, and the boundaries of memory sharing across agents have not yet been made public.
7 more timeline events
All Facts (20)
An Agent's memory mechanism is typically divided into two categories: short-term memory and long-term memory, with long-term memory requiring external persistent storage such as vector databases, relational databases, and graph databases.
80%UnverifiedAgent Memory 的定位是给 AI 编码代理补上长期记忆的本地记忆层,作为所有 Agent 共用的 Memory Server
95%UnverifiedAgent Memory 项目自带一个网页端的实时 Viewer,可以查看记忆的写入、搜索和回放过程
90%UnverifiedAgent Memory 提供插件、Hooks、MCP 协议以及 REST API 等多种接入方式
90%UnverifiedAgent Memory 的记忆数据存储在本地,不依赖云端服务
90%UnverifiedAgent Memory 是一个开源项目,目前已在 GitHub 上获得 10.8k stars
85%UnverifiedAgent Memory 具备语义级别的检索能力,通过向量嵌入将文本转化为高维空间中的数值向量进行语义匹配
85%UnverifiedAgent Memory 的 Readme 中给出了 Long-term Memory Retrieval Benchmark 的数据来展示其语义检索能力
85%Unverified在Agent Memory领域社区讨论中,强化学习与知识内化方向获得最多支持,被认为是Memory的终极目标
75%UnverifiedA memory system should not be merely a passive storage-and-retrieval pipeline, but can actively construct associative structures among memories and endow memories with flexible expressive capability.
70%Unverified几乎所有Agent Memory框架都围绕CRUD四个操作展开,通过embedding做向量检索,方法高度雷同
70%UnverifiedAgent Memory 支持 Claude Code、Codex、Cursor、Gemini CLI 等多种 AI 编码工具接入同一套记忆系统
60%UnverifiedCurrent engineering practice generally adopts a "frozen parameters + external memory read/write" architecture to manage Agent memory, rather than updating weights at inference time.
50%UnverifiedAnthropic calls the compression process in Agent Memory that turns raw conversation streams into usable context "dreaming."
50%UnverifiedFor developers, practicable approaches include: writing key facts into structured notes rather than relying on model memory; setting clear read and write rules for the agent to limit what it can modify; and actively verifying whether memory has been loaded correctly when switching sessions.
50%UnverifiedMemory loading is automatic, and users do not need to manually remind the AI what was discussed last time.
50%UnverifiedAutomatic loading means the model must make trade-offs within limited context, and the quality of the retrieval strategy directly determines the accuracy of memory.
50%UnverifiedThe system can set rules for the agent, requiring it to fill in missing necessary notes or maintain a work record in a prescribed format, which is called "automatic read control" and "note rules."
50%UnverifiedWell-designed memory requires filtering rules, retaining only feedback with universal value for future tasks and avoiding sedimenting one-off changes into permanent rules, otherwise the memory will bloat and produce conflicts.
50%UnverifiedThe Agent's memory mechanism solves the stateless problem by persisting key information in external files; during operation the Agent writes valuable feedback, preferences, and lessons into the memory file, and reads it first on the next startup.
50%