INPUT / 01
素材变成AI的基础特征Assets become foundational AI features
人设、世界观、台词、图片、语音、运营规则——拆成机器可读的资产单元。Canon, scripts, imagery, voice, constraints and operating rules become machine-readable assets.
IP CanonAsset IndexAudit
HARNESS TECHNOLOGY (BEIJING) CO., LTD.
聊过一次，它就记住你。 One chat. It remembers you.
Harness 建设关系记忆基础设施。从 IP 角色切入，让 AI 从一次性对话变成可运营的长期联系人——记住偏好、尊重边界、主动维护关系。 Harness builds relationship-memory infrastructure. Starting from IP characters, we turn AI from disposable chats into long-term contacts that remember preferences, respect boundaries, and maintain relationships proactively.
01 / 角色入口01 / CHARACTER ENTRY
角色，彻底活过来Characters that truly come alive
IP 角色自带认知——用户知道他是谁，愿意为声音、剧情、关系权益付费。Harness 把角色从静态内容推进为可持续运营的 AI 联系人。 IP characters carry instant recognition. Users know who they are and will pay for voice, story arcs and relationship rights. Harness moves them from static content into operable AI contacts.
CONTRACT / 02
实时活人引擎A living, real-time engine
让角色像真人一样有思考、行为、实时状态变化。Characters think, act, and change state in real time—like real people.
ScopeSafetyVersion
OPERATE / 03
互动即关系运营From interaction to relationship operations
私聊、语音、剧情事件、主动触达——所有触点共享同一份关系状态。Chat, feeds, comments, voice, images, story events and outreach share one relationship state.
DirectorTriggerDelivery
02 / BRAINMESH
BrainMesh 让模型理解关系，不只是语言。BrainMesh helps models understand relationships, not just language.
用户、角色、偏好、边界、共同经历——装配成可计算、可观测的关系状态。模型负责生成，BrainMesh 负责关系真相。 Users, characters, preferences, boundaries and shared history are composed into computable, observable relationship state. Models generate. BrainMesh holds the relationship truth.
记忆对象Memory objects
用户画像、角色合约、关系状态、偏好记忆、边界记忆、触发状态、λ 信号——全部是可治理的对象。Human profiles, character contracts, relationship state, events, preferences, boundaries, trigger state and λ signals become governable objects.
混合检索Hybrid retrieval
metadata filter、BM25、pgvector、GraphRAG 和多模态索引共同管理角色资料、用户记忆、剧情状态与审计证据。Metadata filters, BM25, pgvector, GraphRAG and multimodal indexes manage character assets, user memory, story state and audit evidence.
写入 / 治理 / 读取Write / manage / read
候选记忆写入后，经去重、冲突检测、降权和遗忘，最终按当前任务装配上下文。Candidate memories are written, deduplicated, conflict-checked, down-ranked or forgotten, then assembled into context for the current task.
03 / λ SIGNAL
λ 衡量一次互动转化为长期记忆的效率。λ measures how efficiently one interaction becomes long-term memory.
λ 是关系记忆的转化效率指标。它写入证据账本，驱动贝叶斯更新，决定哪些记忆该保留、哪些该召回、何时主动触达。 λ is the yield rate of relationship memory. It enters the evidence ledger and Bayesian updates, deciding which memories to keep, recall, and when to reach out.
λ
λ = wᵀ M x ∈ ℝ — 向量矩阵乘法，非标量除法λ = wᵀ M x ∈ ℝ — vector-matrix product, not scalar division
向量矩阵定义Vector-matrix definition
λ
=
ww
ᵀ
·
m₁₁m₁₂m₁₃m₁₄m₁₅m₁₆
m₂₁m₂₂m₂₃m₂₄m₂₅m₂₆
m₃₁m₃₂m₃₃m₃₄m₃₅m₃₆
m₄₁m₄₂m₄₃m₄₄m₄₅m₄₆
m₅₁m₅₂m₅₃m₅₄m₅₅m₅₆
m₆₁m₆₂m₆₃m₆₄m₆₅m₆₆
·
xx
维度映射Dimension mapping
w ∈ ℝ⁶ · 质量权重quality weights
x ∈ ℝ⁶ · 成本控制cost controls
M ∈ ℝ⁶ˣ⁶ · 交互矩阵interaction matrix
输出标量，非除法scalar output, no division
w 向量 — 记忆质量维度w vector — memory quality dimensions
w₁ 理解感 · w₂ 复访率 · w₃ 低重复w₁ comprehension · w₂ revisit · w₃ less repetition
w₄ 边界遵守 · w₅ 主动回复 · w₆ 付费意愿w₄ boundary · w₅ proactive reply · w₆ pay willingness
x 向量 — 计算成本维度x vector — computational cost dimensions
x₁ 推理 · x₂ 多模态 · x₃ 审核x₁ inference · x₂ multimodal · x₃ audit
x₄ 延迟 · x₅ 打扰 · x₆ 合规成本x₄ latency · x₅ interruption · x₆ compliance
证据分数Evidence score
当前证据显示：这段记忆值得进入关系级召回，但主动触达需要冷却策略约束。 Current evidence: this memory qualifies for relationship-level recall, but proactive outreach needs cooldown constraints.
04 / HAAH NETWORK
Human-Agent 到 Agent-Human，关系翻转成网络。From Human-Agent to Agent-Human, relationships invert into a network.
HAAH 不只是 Agent 间转发消息。人的授权关系画像跟随 Agent 流动——Agent 先替人完成预沟通、偏好对齐、边界确认和关系维护。 HAAH is more than A2A message passing. Authorized relationship profiles travel with agents—agents first complete pre-communication, preference alignment, boundary confirmation and relationship maintenance on behalf of humans.
05 / BUSINESS PATH
关系创造价值。Relationships create value.
精品角色复访率、记忆正反馈、主动触达回复率、角色一致性评测、λ 标签沉淀。 Premium IP character revisit rates, memory positive feedback, proactive outreach reply rates, character consistency evaluation, accumulated λ labels.
产品层Surface
验证什么What it proves
商业路径Commercial route
精品 IP 角色Premium IP contacts
D1 / D7 revisit, voice, images, story rights
用户愿意为"被记住"付费——被主动想起、被具体回应。Users pay to be remembered, proactively reached and specifically answered.
Subscription / rights / revenue share
内容方运营工厂Character operating factory
import, contract, publish, audit
内容方素材快速变成可验收、可迭代、可运营的 AI 联系人。Partner materials rapidly become verifiable, iterable and operable AI contacts.
B2B tools / IP split / managed launch
关系记忆服务Relationship-memory service
BrainMesh API, RSAAS, vertical social model
关系事件、授权画像、A2A 协作、λ 标签——沉淀为训练与运行时资产。Relationship events, authorized profiles, A2A collaboration and λ labels become training and runtime assets.
API / industry cases / model service
BUILD WITH HARNESS
构筑AI时代的新型社交网络。Building the social network of the AI era.
路径：IP 角色授权试点，关系记忆 API 与行业共建。 Path: IP character licensing pilots, relationship-memory API, and industry co-creation.