Best Open-Source Agent Memory Systems (Self-Hosted, 2026)

Vectorize 2026-08-12 15.1 KB 943 17 𝕏 文章列表

本文深入比较了2026年可自托管的开源AI代理记忆系统，包括Mem0、Cognee、Graphiti/Zep、Letta和Hindsight。文章强调开源与可自托管并非同义，并提出了五个评估维度：检索准确性、自托管现实、功能对等性、许可证和架构。通过公开基准测试和实际部署要求，作者揭示了各系统的真实优劣，特别指出Mem0社区最大但基准测试饱和，Graphiti擅长时间推理但需外部图数据库，Hindsight在准确性和部署简便性上领先。文章旨在帮助读者根据自身需求选择最合适的记忆层系统。

Agent记忆开源自托管 Mem0 Cognee Graphiti Letta Hindsight 基准测试向量数据库

# Best Open-Source Agent Memory Systems (Self-Hosted, 2026) **作者**: Vectorize **日期**: 2026-08-11T18:34:05.000Z **来源**: [https://x.com/Vectorizeio/status/2087246190614356190](https://x.com/Vectorizeio/status/2087246190614356190) --- ![Image](/doc/4233/image/img_205177df0424.jpg) "Open source" and "self-hostable" get used as synonyms, and they are not the same promise. Plenty of agent-memory tools that call themselves open source still expect you to stand up an external vector database, run a graph database next to it, or reach for a credit card at the tier you actually need in production. If you're picking a memory layer you can run yourself, that gap matters more than a star count. Here is the honest version: what each open-source agent memory system really costs to run, how they're architected, and how they score when you measure instead of guess. New to the category? Start with [what agent memory is](https://vectorize.io/what-is-agent-memory). This post is the open-source cut of the [full comparison of all major frameworks](https://vectorize.io/articles/best-ai-agent-memory-systems); here we only cover systems you can self-host. Fair warning on the genre: most "best open-source memory" roundups are published by one of the vendors and, conveniently, rank that vendor first. This one is written by one too. So instead of asserting a winner, it leans on two things you can check yourself: published benchmark numbers and what each system actually requires to run. ## How to compare open-source agent memory Five axes decide this, and only two of them show up on a landing page: 1. Retrieval accuracy — does it return the right memory, measured on a public benchmark, not asserted. See [how to evaluate an agent-memory system](https://hindsight.vectorize.io/blog/2026/07/31/evaluate-agent-memory-system) for the full rubric. 2. Self-hosting reality — can you run it with your own infrastructure, and what does "your own infrastructure" include? One container, or a vector DB plus a graph DB plus a queue? 3. Feature parity when self-hosted — do you get the whole product, or a stripped core that nudges you toward the paid cloud? 4. License — permissive enough to actually deploy. 5. Architecture — vector, graph, or hybrid, and what that implies for the queries you can ask. Weight them for your case. An audit-trail agent cares most about temporal modeling; a lean self-hoster cares most about how many databases land on their ops plate. ## The open-source memory systems at a glance Numbers verified against GitHub and the published benchmarks in August 2026. | System | GitHub ★ | License | Self-host reality | Storage | Strongest at | | --- | --- | --- | --- | --- | --- | | Mem0 | ~63.0K | Apache-2.0 | OSS core; embedded store by default, external vector DB for production; hosted tier | Vector (+ optional graph) | Community + conversational personalization | | Cognee | ~30.0K | Apache-2.0 | Embedded defaults, no mandatory cloud | Graph pipeline (SQLite/LanceDB/Kuzu) | Graph ECL pipelines | | Graphiti / Zep | ~29.8K | Apache-2.0 | Graphiti self-hosts; Zep CE deprecated | Temporal knowledge graph (Neo4j/FalkorDB) | Temporal reasoning | | Letta (ex-MemGPT) | ~24.2K | Apache-2.0 | Self-hostable server | Agent runtime + memory | Full agent runtime | | Hindsight | ~19.6K | MIT | One Docker command, embedded PostgreSQL, no external DB | Single embedded Postgres, multi-strategy | Accuracy + deployment simplicity | Two honest notes before the profiles. Hindsight is not the most-starred project here; Mem0, Cognee, and Graphiti all have larger communities, and that's worth something real. And every system in this table is permissively licensed, so "it's open source" is table stakes, not a differentiator. The differences that matter are accuracy and what it takes to run. ## The systems, one by one Mem0 — the biggest community Mem0 is the most popular agent-memory project on GitHub by a wide margin (~63.0K stars), and popularity compounds: more integrations, more examples, more Stack Overflow answers when you're stuck. It's genuinely good at conversational personalization: memories scope by user_id, agent_id, and run_id, so recalling one person's preferences across sessions is the paved path, and you can start with the library, graduate to a self-hosted server, or use the hosted platform. The write side extracts facts from turns rather than storing raw transcript, which keeps the store from ballooning. The trade-offs are less about a night-and-day accuracy gap and more about which benchmark you trust and the deployment shape. Mem0 publishes strong LongMemEval numbers, competitive with the top of the field, but LongMemEval is close to saturated now: the serious systems all cluster in the 90s, so it barely separates them. The sharper test is BEAM at 10 million tokens, where Hindsight is #1 and Mem0 doesn't publish a comparable result. On deployment, Mem0's open-source core runs with an embedded store by default, though production setups typically move to an external vector database, and the smoothest path nudges toward the hosted platform. If community size and a personalization-first design are your priorities, Mem0 is the obvious pick. Graphiti / Zep — the temporal specialist This is the one to reach for when when a fact was true is the whole job. Graphiti, the open-source engine under Zep, models time natively: every edge in its knowledge graph carries validity metadata, so it answers "what was the customer's plan before the Q2 change?" cleanly. For compliance timelines and audit trails, that temporal depth is hard to match, and it's the axis where Graphiti genuinely leads. Two caveats for self-hosters. Zep Community Edition was deprecated in 2025, so the free, open-source path to "self-host Zep with full features" is gone — what you self-host is Graphiti, the engine, not the whole product (enterprise BYOC aside). And Graphiti typically stores its graph in an external database (Neo4j or FalkorDB, though an embedded FalkorDB lite mode exists), so in most deployments you're running and operating that alongside your app. Letta — the agent runtime Letta (formerly MemGPT) is the odd one out here, in a useful way: it's not really a drop-in memory layer, it's an agent runtime with memory built in. If you want the whole framework — the agent loop, tool use, and memory as one opinionated system — Letta is a strong, self-hostable choice with a real research lineage. Its memory model is worth knowing: agents hold editable memory blocks (a small, always-in-context core the agent rewrites itself) plus archival memory for the long tail, an idea straight from the MemGPT paper. If you already have an agent and just want to bolt on memory, it's heavier than you need, because you're adopting a runtime to get a memory feature. Cognee — the graph-pipeline option Cognee's angle is an ECL (extract-cognify-load) graph pipeline with pluggable, embedded-by-default storage (SQLite, LanceDB, Kuzu), which makes for a genuinely low-friction local start with no mandatory cloud. If you like the idea of a configurable graph-construction pipeline over your data, it's worth a look. The thing to bring your own skepticism to is the ranking. Several of the "best open-source memory" roundups at the top of Google are Cognee's own, and they place Cognee first without independent benchmark numbers. That's the gap this whole post is about: prefer a system that publishes a score on a benchmark you can reproduce over one that ranks itself in a blog it controls. Hindsight — accuracy and the simplest self-host Full disclosure, this is our project, so weigh the claims against the sources rather than the byline. Hindsight's bet is the two axes above: retrieval accuracy and deployment simplicity. On accuracy, it posts the highest published scores among these systems — 94.6% on LongMemEval-s, and #1 on [BEAM](https://hindsight.vectorize.io/blog/2026/04/02/beam-sota), the benchmark that tests memory at 10 million tokens where context stuffing is impossible. Both are on public leaderboards with reproducible methodology, which is the point. It gets there by running [four retrieval strategies in parallel](https://hindsight.vectorize.io/blog/2026/07/31/evaluate-agent-memory-system) (semantic, BM25 keyword, graph traversal, temporal) with cross-encoder reranking, rather than betting on one. On deployment, it's a single Docker command with embedded PostgreSQL — no external vector database, no separate graph database, [no bring-your-own-datastore](https://hindsight.vectorize.io/blog/2026/05/12/case-against-external-vector-dbs-agent-memory) at all. It's MIT-licensed, and the self-hosted build has full feature parity with the managed cloud; nothing is gated. Where it doesn't lead is community size and native temporal-validity modeling, which is exactly why Graphiti is in this list. Also worth knowing LangMem is the sensible default if you're already committed to LangGraph; it's the in-ecosystem option rather than a standalone service. Memobase (~2.8K stars) focuses on user-profile memory and is worth a look for that shape. Memary exists but has been dormant since 2024, so treat it as a reference, not a dependency. ## What "self-hosted" really costs The star counts hide the number that hits your ops budget: how many stateful services you end up running. Group the field by it. - One embedded store: Hindsight keeps everything — memories, entities, graph, vectors — in a single embedded PostgreSQL. Cognee runs embedded too, across three stores (SQLite, LanceDB, Kuzu). Mem0 also ships an embedded default (on-disk Qdrant). - External database for production: Graphiti typically runs on an external graph DB (Neo4j or FalkorDB), and Mem0 deployments usually move to an external vector store at scale. Those are real services to run, back up, and secure. - Cloud pressure: Zep's full product is cloud now that CE is deprecated, and Mem0's smoothest path is its hosted platform. The open-source cores are real, but the paved road often leads off-prem. None of this is disqualifying — a team already running Neo4j pays little marginal cost to add Graphiti. But "self-hosted agent memory" should mean you priced the whole deployment, external dependencies included, and know that [external datastores carry real cost](https://hindsight.vectorize.io/blog/2026/07/28/migrate-agent-memory-off-vector-database) before you wire them in. ## When to pick each - Mem0 — you want the largest community and a personalization-first design. - Graphiti / Zep — your agent's core job is temporal reasoning (audit trails, evolving records) and you're fine running an external graph database. - Letta — you want a full, self-hostable agent runtime, not just a memory layer. - Cognee — you want a configurable graph pipeline with embedded local defaults, and you'll benchmark it yourself. - Hindsight — you want the highest published retrieval accuracy and the simplest deployment (one container, no external DB), and you don't need native fact-validity windows. - LangMem — you're already all-in on LangGraph. ## Frequently asked questions Is Mem0 really open source? Yes, the Mem0 core is Apache-2.0 and self-hostable, and it runs with an embedded store by default. But the smoothest production path is its hosted platform, and at scale most deployments move to an external vector database, so "open source" and "no external services in production" are not the same thing here. Is Zep open source? Partly. Zep's open-source engine is Graphiti (Apache-2.0), which you can self-host. Zep the product is a cloud service, and Zep Community Edition (the self-hostable full product) has been deprecated. If you need the whole Zep feature set on your own infrastructure, that path is gone; what stays open is Graphiti, the engine. Which open-source agent memory is easiest to self-host? By the number of stateful services you have to run, the embedded-storage options win: Hindsight (one Docker container, embedded PostgreSQL, no external database) and Cognee (embedded SQLite/LanceDB/Kuzu defaults). Graphiti adds an external graph database, and Mem0 adds a vector store, both of which you then operate, back up, and secure. Which open-source agent memory is the most accurate? On LongMemEval, most serious systems now score in the 90s (Hindsight is at 94.6% on LongMemEval-s), so that benchmark barely separates them anymore. The sharper test is BEAM, which evaluates memory at 10 million tokens where context stuffing is impossible: Hindsight is #1 there, and most competitors don't publish a comparable result. When a roundup declares a winner, ask which benchmark, at what scale, and with what methodology. Do I need a vector database for agent memory? Not necessarily. Some store their graph in an external database (Graphiti), and some run a separate vector store in production (Mem0, though its default is embedded). Others, like Hindsight, keep everything in a single embedded PostgreSQL. That's why the deployment shape, not just the license, belongs in your comparison. What about LangMem, Memobase, and Memary? LangMem is the in-ecosystem choice if you're already on LangGraph. Memobase (~2.8K stars) is a smaller, user-profile-focused option. Memary has been dormant since 2024, so treat it as a reference rather than something to build on. ## The honest verdict There's no universal winner, only the best fit for what you're optimizing. If you want the biggest community, Mem0. If you're building audit-grade temporal reasoning, Graphiti. If you want the whole agent framework, Letta. If you want a configurable graph pipeline, Cognee. And if you're self-hosting and what you care about is retrieval accuracy plus a boring, single-container deployment you can stand up in a minute and forget, Hindsight is built for exactly that. It self-hosts with one Docker command and embedded PostgreSQL, MIT-licensed, or runs managed on [Hindsight Cloud](https://hindsight.vectorize.io/) with the same feature set. Whichever you choose, choose it on a benchmark you can reproduce and a deployment you've actually priced, not on whose blog ranked whom first. ## 相关链接 - [Vectorize](https://x.com/Vectorizeio) - [@Vectorizeio](https://x.com/Vectorizeio) - [293](https://x.com/Vectorizeio/status/2087246190614356190/analytics) - [what agent memory is](https://vectorize.io/what-is-agent-memory) - [full comparison of all major frameworks](https://vectorize.io/articles/best-ai-agent-memory-systems) - [how to evaluate an agent-memory system](https://hindsight.vectorize.io/blog/2026/07/31/evaluate-agent-memory-system) - [BEAM](https://hindsight.vectorize.io/blog/2026/04/02/beam-sota) - [four retrieval strategies in parallel](https://hindsight.vectorize.io/blog/2026/07/31/evaluate-agent-memory-system) - [no bring-your-own-datastore](https://hindsight.vectorize.io/blog/2026/05/12/case-against-external-vector-dbs-agent-memory) - [external datastores carry real cost](https://hindsight.vectorize.io/blog/2026/07/28/migrate-agent-memory-off-vector-database) - [Hindsight Cloud](https://hindsight.vectorize.io/) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [2:34 AM · Aug 12, 2026](https://x.com/Vectorizeio/status/2087246190614356190) - [293 Views](https://x.com/Vectorizeio/status/2087246190614356190/analytics) --- *导出时间: 2026/8/12 23:44:52* --- ## 中文翻译 # 最佳开源 Agent 记忆系统（自托管，2026） **作者**: Vectorize **日期**: 2026-08-11T18:34:05.000Z **来源**: [https://x.com/Vectorizeio/status/2087246190614356190](https://x.com/Vectorizeio/status/2087246190614356190) --- ![Image](/doc/4233/image/img_205177df0424.jpg) “开源”和“可自托管”经常被当作同义词使用，但它们承诺的并非同一件事。许多自称开源的 Agent 记忆工具，仍然期望你搭建一个外部向量数据库，在旁边运行一个图数据库，或者在生产环境真正需要的层级上掏出信用卡。如果你要选择一个可以自己运行的记忆层，这个差距比星标数更重要。以下是诚实的版本：每个开源 Agent 记忆系统实际运行成本是多少，它们是如何架构的，以及当你实际测量而非猜测时，它们的得分如何。这个类别的新手？从[什么是 Agent 记忆](https://vectorize.io/what-is-agent-memory)开始。这篇文章是[所有主要框架完整比较](https://vectorize.io/articles/best-ai-agent-memory-systems)的开源版本；这里我们只涵盖你可以自行托管的系统。对这个类型先做个公平警告：大多数“最佳开源记忆”的综述都是由某个供应商发布的，并且方便地把自己排在第一。这篇也是由供应商写的。所以，与其断言一个赢家，不如依赖两件你可以自己验证的事情：已发布的基准测试数字，以及每个系统实际运行需要什么。 ## 如何比较开源 Agent 记忆有五个轴决定这一点，其中只有两个会出现在落地页上： 1. 检索准确性——它是否返回正确的记忆，以公开基准测试衡量，而非自我宣称。完整的评估标准见[如何评估 Agent 记忆系统](https://hindsight.vectorize.io/blog/2026/07/31/evaluate-agent-memory-system)。 2. 自托管现实——你能用自己的基础设施运行它吗，“自己的基础设施”包括什么？一个容器，还是一个向量数据库加一个图数据库再加一个队列？ 3. 自托管时的功能对等性——你得到的是完整产品，还是一个被精简的核心，引导你使用付费云服务？ 4. 许可证——足够宽松，可以实际部署。 5. 架构——向量、图或混合，以及这对你可以提出的查询意味着什么。根据你的情况权衡这些。审计追踪 Agent 最关心时间建模；精简自托管者最关心有多少数据库落在他们的运维盘子里。 ## 开源记忆系统一览数字已于 2026 年 8 月针对 GitHub 和已发布的基准测试进行了验证。 | 系统 | GitHub ★ | 许可证 | 自托管现实 | 存储 | 最强项 | | --- | --- | --- | --- | --- | --- | | Mem0 | ~63.0K | Apache-2.0 | 开源核心；默认嵌入式存储，生产环境使用外部向量数据库；托管层级 | 向量（+ 可选图） | 社区 + 对话个性化 | | Cognee | ~30.0K | Apache-2.0 | 嵌入式默认，无强制云依赖 | 图管道（SQLite/LanceDB/Kuzu） | 图 ECL 管道 | | Graphiti / Zep | ~29.8K | Apache-2.0 | Graphiti 可自托管；Zep CE 已弃用 | 时间知识图谱（Neo4j/FalkorDB） | 时间推理 | | Letta（前 MemGPT） | ~24.2K | Apache-2.0 | 可自托管服务器 | Agent 运行时 + 记忆 | 完整 Agent 运行时 | | Hindsight | ~19.6K | MIT | 一条 Docker 命令，嵌入式 PostgreSQL，无外部数据库 | 单一嵌入式 Postgres，多策略 | 准确性 + 部署简单性 | 在介绍之前有两点诚实的说明。Hindsight 不是这里星标最多的项目；Mem0、Cognee 和 Graphiti 都有更大的社区，这确实有价值。而且这个表格中的每个系统都是宽松许可的，所以“它是开源的”只是入场券，不是差异化因素。真正重要的区别在于准确性和运行所需的条件。 ## 逐一介绍各个系统 Mem0——最大的社区 Mem0 是 GitHub 上最受欢迎的 Agent 记忆项目，遥遥领先（约 63.0K 星标），而人气会带来复利效应：更多集成、更多示例、遇到问题时更多 Stack Overflow 答案。它确实擅长对话个性化：记忆按 user_id、agent_id 和 run_id 进行范围划分，因此跨会话回忆一个人的偏好是铺好的路径，你可以从库开始，升级到自托管服务器，或使用托管平台。写入端从对话轮次中提取事实，而不是存储原始记录，这防止了存储库膨胀。权衡之处与其说是天壤之别的准确性差距，不如说更关乎你信任哪个基准测试以及部署形态。Mem0 发布了强劲的 LongMemEval 数据，与顶尖水平相当，但 LongMemEval 现在已接近饱和：严肃的系统都聚集在 90 多分，几乎无法区分它们。更尖锐的测试是 1000 万 token 的 BEAM，Hindsight 排名第一，而 Mem0 没有发布可比较的结果。在部署方面，Mem0 的开源核心默认使用嵌入式存储运行，但生产环境通常迁移到外部向量数据库，而最顺畅的路径则倾向于托管平台。如果社区规模和个人化优先的设计是你的首要考虑，Mem0 是显而易见的选择。 Graphiti / Zep——时间专家当“某个事实何时为真”就是全部工作时，这就是你要找的。Graphiti 是 Zep 下的开源引擎，原生建模时间：其知识图谱中的每条边都携带有效性元数据，因此它能干净地回答“Q2 变更之前客户的计划是什么？”。对于合规时间线和审计追踪，这种时间深度很难被超越，这也是 Graphiti 真正领先的轴。自托管者有两个注意事项。Zep 社区版已于 2025 年弃用，因此“以完整功能自托管 Zep”的免费开源路径已经消失——你能自托管的是 Graphiti 引擎，而不是整个产品（企业 BYOC 除外）。而且 Graphiti 通常将图存储在外部数据库中（Neo4j 或 FalkorDB，尽管存在嵌入式 FalkorDB lite 模式），因此在大多数部署中，你需要与应用程序一起运行和运维它。 Letta——Agent 运行时 Letta（前 MemGPT）在这里是个异类，但以有用的方式：它不是一个即插即用的记忆层，而是一个内置记忆的 Agent 运行时。如果你想要整个框架——Agent 循环、工具使用和记忆作为一个统一的系统——Letta 是一个强大的、可自托管的选择，具有真实的研究血统。它的记忆模型值得了解：Agent 持有可编辑的记忆块（一个小的、始终在上下文中的核心，Agent 自己重写）加上用于长尾的归档记忆，这个想法直接来自 MemGPT 论文。如果你已经有一个 Agent，只想加上记忆功能，它比你需要的更重，因为你为了获得记忆功能而采用了一个运行时。 Cognee——图管道选项 Cognee 的角度是 ECL（提取-认知-加载）图管道，具有可插拔的、默认嵌入式的存储（SQLite、LanceDB、Kuzu），这使得本地启动真正低摩擦，无需强制云依赖。如果你喜欢在你的数据上构建可配置的图构建管道的想法，值得一看。需要你自己保持怀疑的是排名。Google 顶部的一些“最佳开源记忆”综述是 Cognee 自己发布的，它们在没有独立基准测试数据的情况下将 Cognee 排在第一。这正是这篇文章要讨论的差距：优先选择在你可以复现的基准测试上发布分数的系统，而不是在自己控制的博客中自我排名的系统。 Hindsight——准确性和最简单的自托管完全披露，这是我们的项目，所以请根据来源而非作者来权衡这些声明。Hindsight 的赌注是上述两个轴：检索准确性和部署简单性。在准确性方面，它发布了这些系统中最高分——LongMemEval-s 上 94.6%，以及在 [BEAM](https://hindsight.vectorize.io/blog/2026/04/02/beam-sota) 上排名第一，该基准测试在 1000 万 token 下测试记忆，上下文填充不可能实现。两者都在公开排行榜上，方法论可复现，这正是关键。它通过[并行运行四种检索策略](https://hindsight.vectorize.io/blog/2026/07/31/evaluate-agent-memory-system)（语义、BM25 关键词、图遍历、时间）并使用交叉编码器重排序来实现这一点，而不是押注于单一策略。在部署方面，它是一条 Docker 命令，带嵌入式 PostgreSQL——没有外部向量数据库，没有单独的图数据库，[完全不需要自带数据存储](https://hindsight.vectorize.io/blog/2026/05/12/case-against-external-vector-dbs-agent-memory)。它采用 MIT 许可证，自托管版本与托管云具有完全的功能对等性；没有任何功能被锁定。它不领先的地方是社区规模和原生时间有效性建模，这正是 Graphiti 出现在这个列表中的原因。其他值得了解的如果你已经致力于 LangGraph，LangMem 是明智的默认选择；它是生态系统内的选项，而非独立服务。Memobase（约 2.8K 星标）专注于用户画像记忆，值得一看。Memary 存在但自 2024 年以来一直处于休眠状态，所以把它当作参考，而不是依赖。 ## “自托管”的真正成本星标数掩盖了影响你运维预算的数字：你最终运行了多少个有状态服务。按此对领域进行分组。 - 单一嵌入式存储：Hindsight 将所有内容——记忆、实体、图、向量——保存在一个嵌入式 PostgreSQL 中。Cognee 也以嵌入式方式运行，跨三个存储（SQLite、LanceDB、Kuzu）。Mem0 也提供嵌入式默认（磁盘上的 Qdrant）。 - 生产环境的外部数据库：Graphiti 通常在外部图数据库（Neo4j 或 FalkorDB）上运行，Mem0 部署在规模扩大时通常迁移到外部向量存储。这些是需要运行、备份和保护的真实服务。 - 云压力：Zep 的完整产品现在是云服务，因为 CE 已弃用，而 Mem0 最顺畅的路径是其托管平台。开源核心是真实的，但铺好的路往往通向外部。这些都不是否定性的——已经在运行 Neo4j 的团队添加 Graphiti 的边际成本很小。但“自托管 Agent 记忆”应该意味着你已经为整个部署定价，包括外部依赖，并且在接入之前知道[外部数据存储会带来真实成本](https://hindsight.vectorize.io/blog/2026/07/28/migrate-agent-memory-off-vector-database)。 ## 何时选择哪个 - Mem0——你想要最大的社区和个人化优先的设计。 - Graphiti / Zep——你的 Agent 核心工作是时间推理（审计追踪、演化记录），并且你愿意运行外部图数据库。 - Letta——你想要一个完整的、可自托管的 Agent 运行时，而不仅仅是记忆层。 - Cognee——你想要一个可配置的图管道，带嵌入式本地默认，并且你会自己进行基准测试。 - Hindsight——你想要最高的已发布检索准确性和最简单的部署（一个容器，无外部数据库），并且你不需要原生事实有效性窗口。 - LangMem——你已经完全投入 LangGraph。 ## 常见问题 Mem0 真的是开源的吗？是的，Mem0 核心是 Apache-2.0 且可自托管，默认使用嵌入式存储运行。但最顺畅的生产路径是其托管平台，而且在规模扩大时大多数部署会迁移到外部向量数据库，所以“开源”和“生产环境中无外部服务”在这里不是一回事。 Zep 是开源的吗？部分是。Zep 的开源引擎是 Graphiti（Apache-2.0），你可以自行托管。Zep 产品本身是云服务，Zep 社区版（可自托管的完整产品）已被弃用。如果你需要在自己的基础设施上获得完整的 Zep 功能集，这条路已经关闭；保持开放的是 Graphiti 引擎。哪个开源 Agent 记忆最容易自托管？按你必须运行的有状态服务数量，嵌入式存储选项胜出：Hindsight（一个 Docker 容器，嵌入式 PostgreSQL，无外部数据库）和 Cognee（嵌入式 SQLite/LanceDB/Kuzu 默认）。Graphiti 增加了外部图数据库，Mem0 增加了向量存储，两者都需要你运维、备份和保护。哪个开源 Agent 记忆最准确？在 LongMemEval 上，大多数严肃系统现在都得分在 90 多分（Hindsight 在 LongMemEval-s 上为 94.6%），所以这个基准测试已经几乎无法区分它们。更尖锐的测试是 BEAM，它在 1000 万 token 下评估记忆，上下文填充不可能实现：Hindsight 在那里排名第一，而大多数竞争对手没有发布可比较的结果。当一篇综述宣布赢家时，问问是哪个基准测试、什么规模、什么方法论。我需要为 Agent 记忆使用向量数据库吗？不一定。有些将图存储在外部数据库中（Graphiti），有些在生产环境中运行单独的向量存储（Mem0，尽管其默认是嵌入式的）。其他如 Hindsight 将所有内容保存在一个嵌入式 PostgreSQL 中。这就是为什么部署形态，而不仅仅是许可证，应该纳入你的比较。 LangMem、Memobase 和 Memary 怎么样？LangMem 是如果你已经在 LangGraph 上的生态系统内选择。Memobase（约 2.8K 星标）是一个较小的、专注于用户画像的选项。Memary 自 2024 年以来一直处于休眠状态，所以把它当作参考，而不是可以构建的东西。 ## 诚实的结论没有普遍的赢家，只有最适合你优化目标的选择。如果你想要最大的社区，选 Mem0。如果你在构建审计级时间推理，选 Graphiti。如果你想要完整的 Agent 框架，选 Letta。如果你想要可配置的图管道，选 Cognee。如果你在自托管，并且关心的是检索准确性加上一个无聊的、单容器部署，可以在一分钟内搭建好然后忘记它，Hindsight 正是为此而生的。它通过一条 Docker 命令和嵌入式 PostgreSQL 自托管，MIT 许可证，或者在 [Hindsight Cloud](https://hindsight.vectorize.io/) 上以相同功能集托管运行。无论你选择哪个，都要基于你可以复现的基准测试和你实际定价过的部署来做决定，而不是基于谁的博客把谁排在第一。 ## 相关链接 - [Vectorize](https://x.com/Vectorizeio) - [@Vectorizeio](https://x.com/Vectorizeio) - [293](https://x.com/Vectorizeio/status/2087246190614356190/analytics) - [what agent memory is](https://vectorize.io/what-is-agent-memory) - [full comparison of all major frameworks](https://vectorize.io/articles/best-ai-agent-memory-systems) - [how to evaluate an agent-memory system](https://hindsight.vectorize.io/blog/2026/07/31/evaluate-agent-memory-system) - [BEAM](https://hindsight.vectorize.io/blog/2026/04/02/beam-sota) - [four retrieval strategies in parallel](https://hindsight.vectorize.io/blog/2026/07/31/evaluate-agent-memory-system) - [no bring-your-own-datastore](https://hindsight.vectorize.io/blog/2026/05/12/case-against-external-vector-dbs-agent-memory) - [external datastores carry real cost](https://hindsight.vectorize.io/blog/2026/07/28/migrate-agent-memory-off-vector-database) - [Hindsight Cloud](https://hindsight.vectorize.io/) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [2:34 AM · Aug 12, 2026](https://x.com/Vectorizeio/status/2087246190614356190) - [293 Views](https://x.com/Vectorizeio/status/2087246190614356190/analytics) --- *导出时间: 2026/8/12 23:44:52*