VecminDB 产品说明书

在多智能体（Multi-Agent）与群体智能时代，构建具备持久学习、自我进化和安全协同能力的智能体大脑已成为行业核心壁垒。传统的向量数据库仍停留在“冷数据索引与静态召回”的旧时代，无法解决“记忆无限膨胀、检索噪音退化、主权隐私外溢、跨租户经验孤立”的痛点。

VecminDB 是全球首款面向多智能体架构设计的认知型向量数据库 (Memory OS)。我们基于“记忆即模型 (MKaM)”哲学，让智能体的记忆具备自演化、自提纯的参数记忆流形 (PMM)能力。

1. 核心代际技术特征

1.1 长短期记忆生命周期自演化 (LTSM Pipeline)

VecminDB 模拟生物大脑的分级记忆模式：

  • 工作记忆 (Working Memory)：在独立的 Column Family 与 WAL 中进行极速的写入和近实时索引；
  • 情节记忆 (Episodic Memory)：通过评估访问频次与语义密度的 Promotion 机制，异步归档到主分片存储中；
  • 参数记忆流形 (PMM)：后台 Vacuum 任务利用 PCA 协方差自适应分解，将失效的情节记忆冷凝提纯为抽象质心 (Abstract Centroid)，并物理删除原始碎片向量，控制存储容量呈常数级收敛。
  • 语义剪枝 (Semantic Pruning)：自动检测并清理超过 90 天未被智能体访问/召回的僵尸质心，从物理上防止语义噪声污染。

1.2 主权联邦认知网络 (Sovereign Federation)

多智能体在分布式局域网内进行协作进化的安全底盘：

  • 主权边界隔离：每个 Tenant 拥有唯一 `Sovereignty Token`，原始记忆保留在本地；
  • 本地实体网关 (Entity-Sensitive Gating)：实时一票否决含有敏感 PII 正则特征的数据外流；
  • 异常值裁剪哨兵 (Outlier Pruning Sentinel)：基于 3-Sigma 统计分步规则，自动拦截和清除恶意注入的对抗性噪声向量；
  • 差分隐私联邦提纯 (DP-Federated PCA)：在全局累加的协方差矩阵中注入 Laplace 噪声，并向第一主成分叠加 10% 的方向偏置，使得联盟质心完全不可逆向逆推。

1.3 索引算法自适应热寻优 (Shadow Indexing & NSGA-II)

  • 无感热切换影子索引 (Shadow Indexing)：当后台重建索引时，系统采用 A/B 双指针路由和双写对齐机制，让智能体在索引重建期间依然享受零延迟的读写服务。
  • 多目标遗传参数调优 (NSGA-II)：运行时利用遗传算法在 4 维目标（ recall、latency、memory、build_time）中寻找 Pareto 最优解，自动计算最契合当前智能体访问模式的索引拓扑（如 HNSW 的 `ef/M` 参数）。

VecminDB Product Specification

In the era of Multi-Agent and swarm intelligence, building autonomous agents with perpetual learning and safe cooperation is a critical competitive barrier. Traditional vector databases only provide static storage and raw retrieval, failing to handle memory bloat, privacy leakage, and knowledge isolation.

VecminDB is the world's first **Memory OS (Cognitive Vector Database)** built for multi-agent systems, following the **Memory as a Model (MKaM)** philosophy to condense sparse vectors into autonomous **Parameter Memory Manifolds (PMM)**.

1. Generative Technology Features

1.1 Long-Term Semantic Memory (LTSM Pipeline)

VecminDB models biological brains by categorizing agent memories into levels:

  • Working Memory: High-frequency ingestion and near-realtime indexing in isolated Column Families and WALs.
  • Episodic Memory: Promoted into shards dynamically based on recall frequency and semantic value.
  • Parameter Memory Manifolds (PMM): Background Vacuum processes compress stale episodic memories into compact **Abstract Centroids** using PCA covariance decomposition.
  • Semantic Pruning: Zombie centroids that have not been accessed/recalled for 90 days are dynamically purged to avoid semantic noise.

1.2 Sovereign Federation Cognitive Network

Safe Swarm Intelligence for distributed multi-tenant collaboration:

  • Sovereign Isolation: Isolated via unique `Sovereignty Tokens` preserving local raw vectors.
  • Entity-Sensitive Gating: Realtime PII regex filtering blocking outbound private data.
  • Outlier Pruning Sentinel: Detects and evicts adversarial or drifting noise vectors using a dynamic 3-Sigma cosine filter.
  • DP-Federated PCA: Injects Laplace differential privacy noise and a 10% principal component directional bias to guarantee absolute mathematical irreversibility.

1.3 Adaptive Index Hyperspace Auto-Tuning

  • Shadow Indexing: Uses dual pointer A/B swapping and double-writing to rebuild index topologies with zero-downtime.
  • Multi-Objective Tuning (NSGA-II): Explores the 4-dimensional objective space (recall, latency, memory, build_time) to locate the Pareto front and adjust HNSW parameters dynamically.