Humalike

• 1. overview (477 chars)

  Humalike is a "behavioral infrastructure" layer for AI agents. It was launched by the startup Soofte Inc in 2025. It aims to complement large-model dialogue agents with social intelligence-capabilities such as turn-taking, normative learning, theory of mind, and social memory. It does not bind a basic model, but is superimposed on existing agents in a model- and scene-independent manner. It focuses on solving the experience gap of "AI is not like humans" in multi-person asynchronous group chats. It is currently open to developers in a free value-added model.

• 2. Product background (607 chars)

  Humalike's operating entity is Soofte Inc, founded in 2025 by Martí Carmona Serrat, Mateusz Jacniacki and others. The team’s judgment is that most current LLM-based conversational agents are technically adequate, but are still socially clumsy: unable to grasp the rhythm of conversation and unable to read the subtle clues that define human interaction. What Humalike wants to do is not to make another chatbot, but to break down the "human-like" thing into callable behavioral components and sell them to developers, CTOs and ML engineers who need real social experience. There is currently no financing scale and revenue data in the public information, and the products are mainly self-service access by developers.

• 3. Product functions and usage experience (1386 chars)

  Humalike encapsulates social capabilities into seven types of composable behavioral components and connects to the existing agent stack through a unified API layer. The core is the Turn-Taking API, which intelligently determines when an agent should speak, wait or interrupt based on social norms and memory, and manages the rhythm of the conversation.   The second step is norm detection and adaptation, allowing the agent to learn the group’s internal jokes, hidden rules, tone of voice, and conversational style, and behave like an insider rather than an outsider. Theory of Mind tracking continues to maintain an evolving view of each participant's "what they believe, what they want, and how they feel" as the conversation progresses, supporting a smoother and more situationally aware response.   Social memory allows agents to remember individual and group norms, opinions, and behavioral patterns across multiple conversations to create continuity; social observability helps agents sense their own and others’ engagement and identify who is participating and who is distracted; social signals are responsible for reading subtle human clues that standard agents often ignore, such as typing pauses, message editing, and deletion reactions, so as to more accurately interpret true intentions. The character (Persona) provides an anchor for the basic personality, allowing the output to remain solid and point-of-view, rather than falling into general AI rhetoric. It is worth emphasizing that these components explicitly support both one-on-one and group multi-user settings, rather than just serving a single round of Q&A.

• 4. Pricing and business model (587 chars)

  Humalike operates on a freemium model. New accounts can get 20 US dollars in free credits, no credit card required, which is enough to connect the API to the agent for testing and prototype verification. The finer payment tiers and pay-as-you-go billing rules have not yet been fully disclosed on the public page, and some advanced behavioral capabilities are still marked as "coming soon." For highly regulated corporate buyers, this means the true cost is confirmed to the sale once the credits are exhausted. The product also provides a one-time Hermes plug-in (GitHub repository Humalike/hermes-humalike-plugin) to facilitate refactoring-free access for teams using Nous Research Hermes.

• 5. User reviews (777 chars)

  As an infrastructure product that only started in 2025, Humalike's real end-user reviews are still relatively sparse. The reviews mainly come from editorial descriptions in developer directories and tool aggregation sites. Positive reviews focus on two points: First, it truly targets the long-term pain point of "AI does not look like humans in multi-user group chats", rather than reinventing the chat shell; second, the model-independent and stack-independent design lowers the access threshold, and developers can use it as a layer of "behavioral middleware" stacked on existing systems.   Negative or reserved opinions point to maturity - some features are still marked as coming soon, the actual effect of behavioral capabilities is highly dependent on the quality and quantity of community, domain and interaction data, and the documentation also admits that performance will vary depending on the scenario.

• 6. Industry evaluation (810 chars)

  In the aibase product library, aipure, utilo, TheAI and other Chinese and English tool catalogs, Humalike is mostly classified into the "AI agent/virtual digital human/multi-user agent" track, labeled as free value-added, and the most suitable scenarios are generally pointed to AI game characters and NPCs. From an industry perspective, it is generally believed that injecting social intelligence into agents is an underestimated but urgently needed direction, especially games, educational technology, community robots and companionship experiences that lack this layer.   Its differentiation lies in the explicit decomposition of "specifications, roles, theory of mind, social memory, signals, and observability" into API components instead of black box encapsulation. In terms of alternative tools, role-playing chatbots such as Emochi are often mentioned side by side in the directory, but the positioning of the two is not exactly the same.

• 7. Controversies and Risks (734 chars)

  There is currently no major controversy surrounding Humalike, but there are several types of inherent risks worth being aware of. First, compliance certification is still in progress - SOC 2 Type II and ISO 27001 are both marked as "in progress" rather than completed, which is a hard threshold for buyers with strong regulations such as finance and medical care. Second, social intelligence is context-sensitive, and model performance will fluctuate with the quality of interaction data, which may cause misunderstanding or offense in some communities.   Third, long-term social memory involves the retention of individual and group behavioral portraits, and privacy and data governance need to be controlled by developers themselves. Fourth, some functions have not yet been fully opened, and the promise of "coming soon" involves uncertainty in the delivery pace.

• 8. Applicable groups and usage suggestions (794 chars)

  Humalike is most suitable for three types of people: game studios (to be NPCs and AI teammates who can remember players and understand group banter), educational technology platforms (to be classroom assistants that track the status of each learner and have an appropriate tone), and developers in community and customer service scenarios (to be moderators and onboarding agents who can adapt to local cultural norms). It is not suitable for small teams that only want to be a single-round Q&A assistant, or that have extremely high data compliance requirements and cannot wait for certification to be implemented.   If you want to be a companion long-term relationship agent, social memory and role components would be a good starting point. As an alternative, you can first look at role-playing products such as Emochi, but if you want to "add a layer of behavioral capabilities" to your own agent stack, Humalike's API approach is more suitable.

• 9. Summarize (373 chars)

  Humalike fills in the long-vacant "social layer" for AI agents. The model-independent and composable API design lowers the access threshold. It is a behavioral infrastructure worth trying in gaming, education, and community scenarios. It is recommended to use free points to verify compliance with specifications and group chat performance in real multi-user conversations before deciding whether to move the workload to the paid layer.

No in-depth report yet

• 

  AWhite_X

  —

  看了他们 arXiv 上的那篇 HUMA 论文，97 个人的四人群聊实验，参与者猜出 AI 是 AI 的比例是 55.4%。这个数字其实挺有意思的——硬币概率也就 50%，说明做人设的时候他们已经接近图灵测试的边界了。

---

• 

  bigladybug415

  —

  刚在 Discord 里接上了 Humalike 的 turn-taking API，效果比纯 prompt 调教好太多了。之前 bot 每条消息都回，第三天就被全组静音，现在至少知道什么时候闭嘴了。

---

• 

  BMorrisII645

  —

  免费额度给了 20 刀不用绑卡，这倒是挺实在的。我跑了大概两百多次请求还没花完。不过说实话他们的 pricing 页面信息太少，真正上生产之后成本怎么算完全没概念。

---

• 

  Debra.Sanders_777

  —

  SOC 2 和 ISO 27001 还是 in progress，对于要做 enterprise 客户的人来说这个是硬伤。拿这个去走合规审查肯定被卡。

---

• 

  石贞

  —

  没官方 SDK 这事确实烦，全部得自己写 HTTP 调用。文档倒是不错，error message 也挺清楚的，就是希望早点出 Python 和 Node 的 SDK。

---

• 

  Kimberly_Gray_88

  —

  搞了个 therapy bot 的 demo，用了 theory of mind API。它能从文本里推断用户的情绪状态，比我预想中敏感。不过我还是不太放心让它在没有人类监督的情况下跑，毕竟读错了空气后果还挺严重的。

---

• 

  LRobinsonIII68

  —

  这个方向选得挺准的。AI agent 现在不是不够聪明，是太不会做人了。在群里每条消息都回一遍，比最烦的同事还烦。Humalike 切出来的 turn-taking 和 norms 确实是真痛点。

---

• 

  CColeman_202448

  —

  Product Hunt 上 launch 那天拿了第二名，435 票。关注度是有的，但用户反馈还不多。现在还处在 closed beta with design partners 的阶段，等开放了再说吧。

---

• 

  6PMRXVW3L

  —

  上个月我就在等 voice turn-taking，结果 launch 之后发现只支持 text。他们说要陆续上 voice model，但 timeline 完全没有。做语音 agent 的团队只能先观望了。

---

• 

  Shirley_Smith_20202

  —

  Turn-taking API 延迟很低，实测两百毫秒以内。做实时对话场景完全够用。Social Memory 查一次大概多加五十毫秒，也算可以接受。

---

• 

  Vincent.DiazIII

  —

  Hermes 一键集成真是救星，不用改架构就能接入。我们原本用的是 Hermes 框架，一行代码没改就把 turn-taking 和 persona 接进去了。

---

• 

  redostrich550

  —

  最打动我的反而是他们官网 AI Coworkers 页面的描述。它说 agent 进 Slack 之后要么每条都回很快被静音，要么只被点名才吭声慢慢没人用。这个观察太准了，我们团队的 bot 就死在这两种模式上。

---

• 

  trueکیمیاجعفری_2024

  —

  这些 behavior API 可以拆开单独用这点好评。我先只接了 turn-taking 试试效果，不用一次性把七个全上了。渐进式集成的思路对开发者很友好。

---

• 

  钟星

  —

  说真的当一个 AI 知道什么时候该闭嘴的时候，你就知道这个方向走对了。Humalike 那个最火的梗图就是教 AI when to shut up，虽然好笑但真的是刚需。

---

• 

  ChristinaMiller11

  —

  帮客户做一个社区 bot，用了 norms API 之后 bot 能学会群里的黑话和语气了，不再是那种一板一眼的客服腔。用户体验提升挺明显的。

---

• 

  周悦

  —

  现阶段只适合做 group chat 场景，别拿它做简单的 FAQ bot，加一层行为层完全是杀鸡用牛刀。

---

• 

  何丽

  —

  终于有人做这个了！！AI agent 最缺的就是社交智能，天天在那尬聊。Humalike 的方向感我给满分。

---

• 

  Samuel.Lee31

  —

  Social Memory 设计好自己的 schema 是关键，官方文档给了一些参考但不够电商场景化。第一天我花了一个小时来规划到底要存哪些用户属性。灵活性是好事，但学习曲线也不低。

---

• 

  SPhillipsSr565

  —

  看了一下他们的 GitHub，Hermes 插件才 7 个 star、0 个 fork。社区活跃度还有很大的成长空间，希望后面能起来。

---

• 

  MelissaKing_X

  —

  看了一下 neodrop 那篇深度分析，说得挺到位的——Humalike 现在最缺的不是更大的模型，而是把那几个脏问题（数据隐私、责任边界、信号采集的透明性）写成文档。如果你要做 therapy 或者 care 场景，这些避不开。

---

• 

  AbigailSmithIII

  —

  做教育场景的，student bot 接入后能记住每个学生的偏好和学习进度。那个跨会话的社交记忆确实有用，学生昨天问到哪了它记得清清楚楚，不再每次从头来。

---

• 

  Nicole.Sanchez369

  —

  Persona API 基于真实社区数据这点比较有意思。不是那种随便写个背景故事就完事的 bot personality，而是真的有数据支撑的个性和立场。

---

• 

  Jessica_AllenIII

  —

  Social Signals 能检测 typing pauses 和 deleted reactions 这点很厉害。之前做客服 bot，用户打了又删的内容才是真实需求。但说实话这个功能有点吓人，用户知道 bot 在盯着他们编辑记录吗？

---

• 

  Christine_Sanders520

  —

  社交信号能读打字停顿和删改反应，这点挺细，但也要小心别在敏感社区误读冒犯到人。

---

• 

  WRamos_88

  —

  我们做线上社区版主机器人时最怕「外来者语气」，Humalike 的规范检测能学群组的内部梗和隐藏规则，让机器人慢慢像圈内人而不是官方通告机。接入后社区里被当成机器人的投诉少了一半，这点比任何话术模板都管用。不过规范学偏了也要有人兜底，建议留个审核开关。

---

• 

  M_etaHub

  —

  我们把它用在陪伴型长关系代理上，社交记忆让重复对话有连续性、不像每次重启，体验提升明显。建议先用免费积分在真实多用户对话里压一压规范遵循和群聊表现，再决定要不要上付费层，别光看 demo。

---

• 

  fck4x

  —

  给教育科技课堂助理接了这一层，代理能记住每个学生之前聊过什么，多学生对话的语气和节奏明显顺了。我们拿 LoSoNA 跑了规范遵循的评测，结果比裸 LLM 高出一截，算是目前见过最像样的「社交层」方案。不过 SOC 2 还在进行中，合规材料得等他们补齐。

---

• 

  Nancy_Hicks168193

  —

  做游戏 NPC 的，看到它的社交记忆和群组玩笑识别直接心动了，这层能力自己写要掉多少头发。

---

• 

  星辰_8

  —

  客户支持场景接了心智理论加可观察性，不同客户的沟通风格能跟着调，摩擦少了一些。

---

• 

  姜睿杰

  —

  把社交可观察性接进客服代理后，能实时感知对方是不是走神或烦躁，该放慢就放慢、该收尾就收尾，对话完成率肉眼可见地上去了。原先纯靠意图分类，经常在对面已经不耐烦时还在机械地走流程，现在这层信号确实补上了人的分寸感。

No reviews yet