AI 记忆工具两大阵营：Memory Backends 与 Context Substrates

witcheer 2026-04-17 25.6 KB 96.3K 511 𝕏 文章列表

文章深入分析了 GitHub 上的 AI Agent 记忆工具，将其分为两大阵营：一是“记忆后端”，如 Mem0 和 Supermemory，主要提取事实并存储于向量数据库；二是“上下文基底”，如 OpenClaw，通过人类可读的结构化文件让会话在上下文中累积。作者指出，虽然 Camp 1 主导了市场，但 Camp 2 的架构才是支持连续、多会话工作的未来方向，例如 Zep 已开始从“记忆”转型为“上下文工程”。

# I Went Through Every AI Memory Tool I Could Find. There Are Two Camps. **作者**: witcheer ☯︎ **日期**: 2026-04-15T16:44:14.000Z **来源**: [https://x.com/witcheer/status/2044456778843238689](https://x.com/witcheer/status/2044456778843238689) --- ![Image](/doc/639/image/img_11b5d1c1680d.jpg) there are 450+ repos tagged "agent-memory" on github and 460+ tagged "context-management." me and my agentic best friends went through them. what I expected to find: 40 tools doing roughly the same thing with different APIs. what I actually found: two fundamentally different paradigms, almost no one drawing the line between them, and a category that doesn't have a name yet. I run a 24/7 agent setup on a Mac Mini M4. every session compounds on the last. that setup is the reason I noticed the split, most memory tools couldn't power what I'm doing, and the ones that could weren't being talked about as memory tools at all. here's the map. ## The Two Camps > camp 1: memory backends - these tools extract facts from your conversations, store them in vector databases, and retrieve relevant ones when you ask. automated note-takers. they file things away and pull them back when needed. > camp 2: context substrates - these maintain structured, human-readable context that accumulates across sessions. nothing gets "extracted." the context is the files. your agent reads them, works within them, writes back to them, and the whole thing compounds over time. camp 1 asks: "what should the AI remember?" camp 2 asks: "what context should the AI work inside?" most of the space (and most of the github stars) sit in camp 1. but camp 2 is where the architectures that actually scale to continuous, multi-session, multi-project work are emerging. and the language is starting to shift in that direction. ## Camp 1: The Memory Backends Mem0 — 53.1k stars the category leader by adoption. four operations: add, search, update, delete. extracts facts from conversations, stores them at three levels (user, session, agent), retrieves them via hybrid search. dead simple to integrate. python and typescript SDKs. works with everything. the limitation: memories are flat entries. no relationships between them. every extraction requires an LLM call, so quality depends entirely on how good the extraction prompt is. and once stored, they don't evolve, a fact from january sits next to a fact from april with no notion that one might supersede the other. MemPalace — 46.2k stars local-first verbatim memory. instead of extracting facts, MemPalace stores conversations word-for-word and organises them into wings (entities), rooms (topics), and drawers (original content). searches them with ChromaDB. the benchmark numbers are the highest in the space: 96.6% retrieval recall on LongMemEval using raw semantic search alone, no API calls, no LLM. 98.4% with hybrid pipeline. 99%+ with LLM reranking. the limitation: verbatim storage scales linearly. the more you talk, the bigger it gets. no compression, no synthesis. if your problem is "find the thing I said three weeks ago," this is the best tool. if your problem is "give me the current state of my work across five projects," it's the wrong tool. Supermemory — 21.8k stars positions itself explicitly as "memory is not RAG." the differentiator is temporal awareness, say "I just moved to San Francisco" and it supersedes your old city. expired facts get forgotten automatically. user profiles combine stable facts with recent activity at ~50ms retrieval. connectors for google drive, gmail, notion, onedrive, github. multi-modal across PDFs, images, videos, code. they created their own benchmark framework (MemoryBench) and claim #1 on LongMemEval, LoCoMo, and ConvoMem. most camp 1 tools treat facts as permanent. Supermemory treats them as evolving. that's the closest camp 1 gets to thinking about state, not just storage. Honcho — 2.4k stars smaller but architecturally distinct. Honcho treats both humans and agents as "peers" in a unified model. an async reasoning service runs in the background, deriving psychological insights about each peer from their sessions. it's not just remembering what you said, it's building a model of how you think. PostgreSQL + pgvector required. AGPL-3.0 (restrictive). heavier infrastructure than most. the closest thing in camp 1 to caring about entity evolution rather than just fact storage. the rest of camp 1, briefly: Cognee (15.4k) combines vector search with graph databases for relational reasoning. Memori (13.3k) intercepts LLM API calls to capture execution context, hits 81.95% on LoCoMo using only 4.97% of full-context tokens. AgentScope, MemOS, EverOS, MIRIX, SimpleMem, Memobase, all variations on the same loop. ## What Camp 1 Tools Have In Common every tool above runs the same fundamental loop: conversation happens → system extracts facts or stores content → facts go into a database (vector, graph, or both) → next conversation, relevant facts get retrieved and injected the intelligence is in the extraction and retrieval. the human interacts with the agent. the memory system works behind the scenes. you never touch the memory directly and you trust the system to remember the right things and surface them at the right time. this works. the benchmarks prove it. but it's solving one specific problem: fact recall. "what did I say about X?" "what does the user prefer?" there's a different problem none of these tools address. ## Camp 2: The Context Substrates OpenClaw — 358k stars you know what it is already, but its memory architecture is the part that matters here. plain markdown files: MEMORY.md for long-term storage, daily notes (YYYY-MM-DD.md) for running context, DREAMS.md for consolidation summaries. the line from their docs that defines the philosophy: "the model only 'remembers' what gets saved to disk, there is no hidden state." no vector database. no extraction pipeline. files the agent reads and writes to. the most interesting feature is dreaming: a background process that consolidates daily notes into long-term memory in three phases: - light sleep — screens daily notes, groups nearby lines into coherent chunks - REM — weighted recall promotion, frequently-accessed information becomes "lasting truths" - deep sleep — replay-safe promotion into MEMORY.md, reconciles rather than duplicates only entries passing all threshold gates get promoted: minimum score 0.8, minimum recall count 3, minimum unique queries 3. six weighted signals score every candidate, relevance (0.30), frequency (0.24), query diversity (0.15), recency (0.15), consolidation (0.10), conceptual richness (0.06). this is background consolidation of lived context. the system doesn't decide what's a "fact" but it promotes what keeps coming up as relevant. Zep — 4.4k stars Zep recently rebranded their entire positioning from "memory" to "context engineering." that one move is the strongest market signal in this entire landscape. a funded company with 4.4k stars looked at where the space was going and decided "memory" was the wrong word for what they were building. under the hood, Zep uses a temporal knowledge graph (their Graphiti framework). facts include valid_at and invalid_at timestamps. it extracts relationships automatically and returns pre-formatted context blocks optimised for LLM consumption. sub-200ms retrieval. SOC2 Type 2 and HIPAA compliant. Zep sits between the two camps architecturally, it still extracts and retrieves. but the rebrand is the tell. the company closest to the camp 1 / camp 2 boundary chose camp 2's language. Thoth — 145 stars tiny project, deepest architecture I found in the entire landscape. Thoth builds a personal knowledge graph with 10 entity types connected by 67 typed directional relations. FAISS vector search with one-hop graph expansion before every LLM call. the standout is the dream cycle, a nightly four-phase process: duplicate merging at 0.93+ similarity threshold → description enrichment from conversation context → relationship inference between co-occurring entities → confidence decay on relations older than 90 days three anti-contamination layers prevent cross-entity fact bleed. it's the most sophisticated automated memory refinement I found. it's sitting at 145 stars because it requires you to take the camp 2 thesis seriously enough to set up a knowledge graph for your own context. most people don't. worth watching. TrustGraph — 2.0k stars introduces "Context Cores", portable, versioned bundles that contain domain schemas, knowledge graphs, vector embeddings, evidence sources, and retrieval policies. treats context like code: version it, test it, promote it, roll it back. the framing matters. every camp 1 tool treats memory as a side effect of conversations. TrustGraph treats context as a first-class artifact with identity, versioning, and a lifecycle. you can hand a Context Core to a new agent and it inherits the full operational context. you can fork one for an experiment and merge it back. this is the closest thing in the space to what a packaged, portable unit of context looks like. the implementation is heavy (Cassandra + Qdrant), but the conceptual model is the right one. MemSearch (by Zilliz) — 1.2k stars markdown-first memory from the team behind Milvus. memories are .md files, human-readable, editable, version-controllable. Milvus runs as a "shadow index" derived from the files, fully rebuildable. the files are the source of truth. the vector search is just an access layer on top. three-layer progressive disclosure: semantic chunks → full sections → raw transcripts. hybrid search (dense vectors + BM25 + RRF reranking). what's notable is that this came from Zilliz, a vector database company. they shipped a memory system where their own product is downstream of the files. that's a meaningful concession about where the source of truth actually lives. ## What Camp 2 Tools Have In Common the loop is different: agent reads structured context before working → agent works within that context → agent (or background process) writes back to the structured context → next session, the context is richer than before the intelligence is in accumulation. the context is the memory. and because it's files (markdown, knowledge graphs, context containers), a human can read it, edit it, correct it, and understand exactly what the agent knows. camp 1 optimises for recall: can the system find the right fact? camp 2 optimises for compounding: does the system get better over time? ## Where This Is Heading, And What I'm Working On the pattern from running a 24/7 agent setup is clear. memory and context aren't the same problem. my agent doesn't need to "remember" that I prefer dark mode. it needs to operate inside a context that includes my active projects, the people I work with, recent decisions, and what happened yesterday.. and that context needs to be richer tomorrow than today. memory backends solve recall. 96%+ accuracy, sub-200ms latency, drop-in APIs. if you need a chatbot to remember user preferences, Mem0 or MemPalace will do it. but if you're running an agent continuously, one that works while you sleep, reads from the same knowledge base your other tools write to, and gets meaningfully better over weeks and months, the context substrate approach is what makes that work. my prediction is that within 6 months, "context engineering" replaces "memory" as the default term for what serious agent infrastructure does. the projects building substrate-style architectures will pull ahead of the ones still framing the problem as fact storage. the benchmarks will get rewritten or new ones will replace them. the project I'm working with is ALIVE (alivecontext.com / @AliveContext_). structured context substrate, file-native, agent-agnostic. walnuts as portable context containers. zero infrastructure dependencies, plain files that compound. it's what I run on top of Hermes Agent on the Mac Mini and in Claude Code, and it's the reason that setup actually works instead of resetting every session. the category needs a name. I think it's context substrate. either way, if you're building agents that need to run for more than one conversation, you're going to end up here. ## 相关链接 - [witcheer ☯︎](https://x.com/witcheer) - [@witcheer](https://x.com/witcheer) - [35K](https://x.com/witcheer/status/2044456778843238689/analytics) - [@AliveContext_](https://x.com/AliveContext_) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [12:44 AM · Apr 16, 2026](https://x.com/witcheer/status/2044456778843238689) - [35.9K Views](https://x.com/witcheer/status/2044456778843238689/analytics) - [View quotes](https://x.com/witcheer/status/2044456778843238689/quotes) --- *导出时间: 2026/4/17 00:39:38* --- ## 中文翻译 # 我把能找到的所有 AI 记忆工具都过了一遍。它们分成两大阵营。 **作者**： witcheer ☯︎ **日期**： 2026-04-15T16:44:14.000Z **来源**： [https://x.com/witcheer/status/2044456778843238689](https://x.com/witcheer/status/2044456778843238689) --- ![Image](/doc/639/image/img_11b5d1c1680d.jpg) GitHub 上有 450 多个带 "agent-memory" 标签的仓库，460 多个带 "context-management" 标签的仓库。我和我的智能体挚友们把它们全都过了一遍。我原本以为会发现：40 个做着大致相同事情、只是 API 不同的工具。我实际发现的是：两种根本不同的范式，几乎没有人指出它们之间的分界线，以及一个还没有名字的品类。我在一台 Mac Mini M4 上运行一个 24/7 全天候的智能体系统。每个会话都在上一个会话的基础上累积。正是这套系统让我注意到了这种分化——大多数记忆工具根本支撑不了我在做的事，而能支撑的那些工具，压根没被当作记忆工具来讨论。下面是全景图。 ## 两大阵营 > 阵营 1：记忆后端 —— 这些工具从你的对话中提取事实，存入向量数据库，并在你提问时检索相关内容。自动化的记事员。它们把东西归档，需要时再取出来。 > 阵营 2：上下文基底 —— 这些工具维护结构化的、人类可读的上下文，并跨会话持续累积。没有任何东西被"提取"。上下文就是文件本身。你的智能体读取它们、在其中工作、再写回它们，整个过程随时间不断复利增长。阵营 1 的问题是："AI 应该记住什么？" 阵营 2 的问题是："AI 应该在什么上下文中工作？" 这个领域的大部分（以及大部分 GitHub 星标）都在阵营 1。但阵营 2 才是真正能扩展到持续、多会话、多项目工作的架构正在涌现的地方。行业的语言也开始朝这个方向转变。 ## 阵营 1：记忆后端 Mem0 — 53.1k 星标按采用率来说是品类的领导者。四个操作：添加、搜索、更新、删除。从对话中提取事实，按三个层级（用户、会话、智能体）存储，通过混合搜索进行检索。集成极其简单。提供 Python 和 TypeScript SDK。什么都能配合使用。局限性：记忆是扁平的条目，彼此之间没有关系。每次提取都需要一次 LLM 调用，所以质量完全取决于提取提示词写得好不好。而且一旦存入就不会演化——一月的一条事实和四月的一条事实并排放着，完全没有"前者可能已被后者取代"的概念。 MemPalace — 46.2k 星标本地优先的逐字记忆。MemPalace 不提取事实，而是逐字存储对话内容，并将它们组织成翼（实体）、房间（主题）和抽屉（原始内容）。使用 ChromaDB 进行搜索。它的基准测试数据是该领域最高的：仅靠原始语义搜索，在 LongMemEval 上就达到 96.6% 的检索召回率，无需 API 调用、无需 LLM。使用混合流水线达到 98.4%。加上 LLM 重排序则超过 99%。局限性：逐字存储的规模线性增长。你聊得越多，它就越大。没有压缩，没有综合。如果你的问题是"找出我三周前说过的那件事"，这是最好的工具。如果你的问题是"告诉我我在五个项目上的工作现状"，那就是错误的工具。 Supermemory — 21.8k 星标明确地将自己定位为"记忆不是 RAG"。它的差异化在于时间感知——说一句"我刚搬到旧金山"，它就会取代你的旧城市。过期的事实会被自动遗忘。用户画像将稳定事实与近期活动相结合，检索延迟约 50 毫秒。提供 Google Drive、Gmail、Notion、OneDrive、GitHub 的连接器。跨 PDF、图像、视频、代码的多模态支持。他们创建了自己的基准测试框架（MemoryBench），并声称在 LongMemEval、LoCoMo 和 ConvoMem 上排名第一。大多数阵营 1 工具把事实当作永久不变的。Supermemory 把它们当作不断演化的。这是阵营 1 中最接近"思考状态而非仅仅存储"的做法。 Honcho — 2.4k 星标规模较小，但架构上独树一帜。Honcho 将人类和智能体都视为统一模型中的"对等体"。一个异步推理服务在后台运行，从各个对等体的会话中推导出心理学洞察。它不只是记住你说了什么，而是在构建你如何思考的模型。需要 PostgreSQL + pgvector。AGPL-3.0 许可（限制性较强）。基础设施比大多数工具更重。这是阵营 1 中最接近关注实体演化而非仅仅存储事实的东西。阵营 1 的其他工具，简要一提： Cognee（15.4k）将向量搜索与图数据库结合以实现关系推理。Memori（13.3k）拦截 LLM API 调用来捕获执行上下文，仅使用 4.97% 的全上下文 token 就在 LoCoMo 上达到 81.95%。AgentScope、MemOS、EverOS、MIRIX、SimpleMem、Memobase——都是同一循环的不同变体。 ## 阵营 1 工具的共同点上面提到的每个工具都运行着同一个基本循环：对话发生 → 系统提取事实或存储内容 → 事实进入数据库（向量、图或两者兼有）→ 下一次对话时，相关事实被检索并注入智能体现在提取和检索环节。人类与智能体交互。记忆系统在幕后工作。你从不直接接触记忆，而是信任系统会记住正确的东西并在正确的时机呈现出来。这是有效的。基准测试证明了这一点。但它解决的是一个特定的问题：事实回忆。"关于 X 我说了什么？""用户偏好是什么？" 还有一个不同的问题，这些工具全都没有解决。 ## 阵营 2：上下文基底 OpenClaw — 358k 星标你已经知道它是什么了，但这里重要的是它的记忆架构。纯 Markdown 文件：MEMORY.md 用于长期存储，每日笔记（YYYY-MM-DD.md）用于运行中的上下文，DREAMS.md 用于整合摘要。他们文档中定义这一理念的那句话："模型只'记住'被保存到磁盘的内容，不存在隐藏状态。" 没有向量数据库。没有提取流水线。只有智能体读取和写入的文件。最有趣的功能是做梦：一个后台进程，将日常笔记分三个阶段巩固为长期记忆： - 浅睡——筛选每日笔记，将相邻的行分组成连贯的块 - 快速眼动（REM）——加权回忆晋升，频繁访问的信息变成“持久真相” - 深睡——安全重放式晋升到 MEMORY.md，进行调和而非重复只有通过所有阈值门槛的条目才会被晋升：最低得分 0.8，最低召回次数 3，最低独立查询数 3。六个加权信号为每个候选条目打分：相关性（0.30）、频率（0.24）、查询多样性（0.15）、时效性（0.15）、巩固度（0.10）、概念丰富度（0.06）。这是对生活上下文的后台巩固。系统不决定什么是“事实”，但它会晋升那些不断被证明相关的信息。 Zep — 4.4k 星 Zep 最近将其整个定位从“记忆”重新命名为“上下文工程”。这一举动是整个领域中强烈的市场信号。一家拥有 4.4k 星的融资公司审视了这个领域的发展方向，然后认定“记忆”并不是描述他们所构建内容的正确词汇。在底层，Zep 使用时序知识图谱（他们的 Graphiti 框架）。事实包含 valid_at 和 invalid_at 时间戳。它自动提取关系，并返回为 LLM 消费优化的预格式化上下文块。检索延迟低于 200ms。通过 SOC2 Type 2 和 HIPAA 合规认证。从架构上看，Zep 位于两个阵营之间——它仍然进行提取和检索。但重新命名透露了真相。最接近阵营 1 / 阵营 2 边界的公司选择了阵营 2 的语言。 Thoth — 145 星小项目，却是我在整个领域中找到的最深度的架构。Thoth 构建了一个包含 10 种实体类型的个人知识图谱，实体之间通过 67 种类型化的有向关系连接。在每次 LLM 调用前使用 FAISS 向量搜索加一跳图扩展。最出色的是梦境周期，一个每夜运行的四阶段流程：以 0.93+ 相似度阈值进行重复合并 → 从对话上下文丰富描述 → 推断共同出现实体之间的关系 → 对超过 90 天的关系进行置信度衰减三层防污染机制防止跨实体的事实渗透。这是我发现的最复杂的自动化记忆精炼系统。它只有 145 星，是因为它要求你足够认真地对待阵营 2 的论点，愿意为自己的上下文搭建知识图谱。大多数人不会这么做。值得关注。 TrustGraph — 2.0k 星引入了"Context Cores"（上下文核心），即可移植、带版本控制的打包单元，包含领域模式、知识图谱、向量嵌入、证据来源和检索策略。它把上下文当作代码来对待：版本化、测试、晋升、回滚。这个定位很重要。每个阵营 1 工具都把记忆当作对话的副产品。TrustGraph 则把上下文视为具有身份、版本和生命周期的一等制品。你可以把一个 Context Core 交给新 agent，它会继承完整的运营上下文。你可以 fork 一个用于实验，然后再合并回来。这是该领域中最接近“打包、可移植的上下文单元”应有形态的东西。实现较重（Cassandra + Qdrant），但概念模型是正确的。 MemSearch（Zilliz 出品）— 1.2k 星来自 Milvus 团队的 markdown 优先记忆系统。记忆是 .md 文件，人类可读、可编辑、可版本控制。Milvus 作为从文件派生的“影子索引”运行，完全可重建。文件是事实来源，向量搜索只是其上的访问层。三层渐进式披露：语义块 → 完整章节 → 原始记录。混合搜索（稠密向量 + BM25 + RRF 重排序）。值得注意的是，这来自 Zilliz——一家向量数据库公司。他们发布了一个记忆系统，而他们自己的产品却处于文件的下游。这是对事实来源真正所在之处的重大让步。 ## 阵营 2 工具的共同点循环是不同的： agent 在工作前读取结构化上下文 → agent 在该上下文中工作 → agent（或后台进程）写回结构化上下文 → 下次会话时，上下文比之前更丰富智能在于积累。上下文就是记忆。而且因为它就是文件（markdown、知识图谱、上下文容器），人类可以阅读、编辑、纠正它，并确切理解 agent 知道什么。阵营 1 优化的是召回：系统能否找到正确的事实？阵营 2 优化的是复利：系统是否随着时间变得更好？ ## 未来方向，以及我正在做的事情运营 24/7 agent 配置的模式已经很清晰。记忆和上下文不是同一个问题。我的 agent 不需要“记住”我喜欢深色模式。它需要在一个包含我的活跃项目、我合作的人、最近的决策以及昨天发生了什么的上下文中运行……而这个上下文明天需要比今天更丰富。记忆后端解决的是召回问题。96%+ 的准确率、低于 200ms 的延迟、即插即用的 API。如果你需要一个聊天机器人记住用户偏好，Mem0 或 MemPalace 就能做到。但如果你在持续运行一个 agent——它在你睡觉时工作、从你的其他工具写入的同一知识库中读取、并在数周和数月内显著变得更好——那么上下文基底的方法才是实现这一切的关键。我的预测是，在 6 个月内，“上下文工程”将取代“记忆”，成为描述严肃 agent 基础设施所做工作的默认术语。构建基底式架构的项目将领先于那些仍将问题定位为事实存储的项目。基准测试将被重写，或者被新的基准取代。我正在参与的项目是 ALIVE（alivecontext.com / @AliveContext_）。结构化的上下文基底，文件原生，与代理无关。核桃（walnuts）作为可移植的上下文容器。零基础设施依赖，只是可以不断累积的普通文件。这就是我在 Mac Mini 的 Hermes Agent 和 Claude Code 中运行的东西，也正是那套配置能真正运作起来、而不是每次会话都归零的原因。这个品类需要一个名字。我认为应该叫“上下文基底”（context substrate）。无论如何，如果你在构建需要跨多次对话持续运行的代理，你最终都会走到这一步。 ## 相关链接 - [witcheer ☯︎](https://x.com/witcheer) - [@witcheer](https://x.com/witcheer) - [35K](https://x.com/witcheer/status/2044456778843238689/analytics) - [@AliveContext_](https://x.com/AliveContext_) - [升级到 Premium](https://x.com/i/premium_sign_up) - [2026年4月16日上午12:44](https://x.com/witcheer/status/2044456778843238689) - [35.9K 次浏览](https://x.com/witcheer/status/2044456778843238689/analytics) - [查看引用](https://x.com/witcheer/status/2044456778843238689/quotes) --- *导出时间: 2026/4/17 00:39:38*