Introduction
Project: High-Performance Temporal Vector Analytics Platform in Rust Author: Manuel Couto Pintos
ChronosVector (CVX) is a temporal vector database built in Rust that treats time as a geometric dimension in the embedding space. It doesn’t just store and retrieve vectors — it models the evolution of semantics over time: analyzing drift velocity, detecting change points, projecting trajectories onto clinical or domain-specific anchors, and serving as episodic memory for AI agents.
What Makes CVX Different
Section titled “What Makes CVX Different”Standard vector databases (Qdrant, Milvus, Pinecone) treat vectors as static snapshots. CVX treats them as trajectories — ordered sequences of embeddings that evolve over time. This enables operations that static stores cannot provide:
| Capability | Static Vector DB | CVX |
|---|---|---|
| kNN search | Cosine/L2 at a point in time | Composite semantic + temporal distance |
| Trajectory extraction | Not possible | trajectory(entity_id) — ordered by timestamp |
| Drift measurement | Not possible | velocity(), drift(), hurst_exponent() |
| Change point detection | Not possible | detect_changepoints() (PELT/BOCPD) |
| Anchor projection | Not possible | project_to_anchors() — map to interpretable dimensions |
| Semantic regions | Not possible | regions(), region_assignments() — HNSW hierarchy as clustering |
| Episodic memory | Not possible | Episode encoding, causal search, temporal continuation |
| Path signatures | Not possible | Reparametrization-invariant trajectory descriptors |
The Six Layers
Section titled “The Six Layers”CVX implements a unified theory across six layers:
- Temporal vector index — ST-HNSW with time-decay edges, SIMD distance kernels, RoaringBitmap filtering, centering for anisotropy correction
- Differential calculus — Velocity, drift, Hurst exponent, change point detection (PELT/BOCPD) on embedding trajectories
- Algebraic invariants — Path signatures (reparametrization-invariant), persistent homology, distributional distances (Fisher-Rao, Wasserstein)
- Temporal causality — Episode encoding, causal search (temporal edges), typed edges (success/failure attribution), Granger causality
- Probabilistic reasoning — Bayesian networks (
cvx-bayes) for P(success | context), Region MDP, Bayesian multi-factor scoring with online weight learning - Structural knowledge — Knowledge graph (
cvx-graph) for task plans, shared sub-plans, compositional reasoning, constraint validation
Validated Applications
Section titled “Validated Applications”| Domain | Dataset | Key Result |
|---|---|---|
| Mental health detection | eRisk (1.36M Reddit posts, 2,285 users) | F1=0.744 with DSM-5 anchor projection |
| Political discourse | ParlaMint-ES (32K speeches, 841 MPs) | F1=0.94 predicting speaker gender from rhetoric |
| AI agent memory | HumanEval, ALFWorld, APPS | 6x improvement in task completion (E3) |
| Embedding anisotropy | MentalRoBERTa D=768 | 30x discriminative signal improvement via centering |
Who is this for?
Section titled “Who is this for?”| Persona | Use Case |
|---|---|
| Clinical NLP Researcher | Track symptom drift in social media using DSM-5 anchors |
| Political Scientist | Analyze rhetorical evolution and polarization over time |
| AI/Agent Researcher | Build episodic memory for LLM agents — store and retrieve action sequences |
| ML Engineer | Monitor embedding drift in production models |
| NLP Researcher | Study diachronic semantic evolution with path signatures |
What CVX is NOT
Section titled “What CVX is NOT”- Not a general-purpose vector database. If you don’t need temporal queries, use Qdrant.
- Not a streaming platform. CVX receives vectors, it doesn’t produce them.
- Not a model training framework. Neural ODE is trained externally (TorchScript); CVX is inference + storage + analytics.
- Not distributed. Single-node architecture. Designed for research-scale datasets (up to ~10M vectors).
Tech Stack
Section titled “Tech Stack”- Language: Rust (edition 2024), with Python bindings via PyO3
- Index: HNSW with temporal extensions, SIMD distance kernels (AVX2/NEON via
pulp) - Storage: RocksDB (hot tier), file-based partitions (warm tier), postcard serialization
- Analytics: 20+ modules in
cvx-analytics— calculus, signatures, topology, ODE, Granger, Procrustes - Reasoning:
cvx-bayes(Bayesian networks),cvx-graph(knowledge graphs) - API: Python bindings (primary), REST (axum), gRPC (tonic), MCP server for LLM integration