Skip to content

Aether SDK Roadmap

The Aether SDK is evolving from a research prototype into a robust, production-ready infrastructure for AI context engineering. Our mission is to provide developers with a sovereign, high-performance memory layer that bridge the gap between simple text retrieval and human-like cognitive memory.

🗺️ Product Milestones

Phase 1: Core Foundation

Goal: Establish a robust protocol-driven architecture for vendor-neutral integrations.

  • Finalize Protocol Layer: Ensure clean, stable interfaces for LLMs, Vector Stores, and Knowledge Graphs.
  • Agnostic Adapters: Implement reference adapters for OpenAI, Anthropic, Qdrant, Chroma, and Neo4j.
  • Sovereign Mode: Enable strictly local execution using Ollama and local storage for air-gapped security.

Phase 2: Knowledge Ingestion & Structuring

Goal: Transform raw data into a structured knowledge base with high semantic density.

  • Advanced Parsing: Integrate industry-standard parsers (e.g., docling) for complex document types (PDF, XLSX, etc.).
  • Autonomous Graph Extraction: Implement automatic entity and relationship extraction to build a persistent Knowledge Graph.
  • Entity Resolution: Build intelligent mechanisms to merge duplicate entities and maintain graph integrity.

Phase 3: Hybrid Retrieval & Context Optimization

Goal: Deliver the most accurate context with minimal token overhead.

  • Hybrid Retrieval Engine: Combine semantic vector search with graph traversal (inspired by HippoRAG) for multi-hop reasoning.
  • Context Distillation: Implement active context window optimization to reduce token costs while preserving information density.
  • Recursive Consolidation: Develop background "sleep cycles" to consolidate episodic memories into long-term structured knowledge.

🛠 Features & Capabilities

1. Protocol & Architecture

  • LLM Integration: Standardized interface for any completion provider.
  • Storage Abstraction: Unified CRUD operations for both Vector and Graph databases.
  • Resource Orchestration: High-level handlers for memory lifecycle management.

2. Intelligent Ingestion

  • Multi-Format Parsing: Support for a wide range of data sources and document formats.
  • Semantic Atomization: Converting narrative context into atomic, indexable facts.
  • Schema Evolution: Handling changes in data structure without losing historical context.

3. Advanced Memory Operations

  • Vector search (Episodic): Local and cloud-scale vector retrieval.
  • Graph Traversal (Semantic): Relational reasoning over extracted knowledge.
  • High-Throughput Buffer: Asynchronous ingestion for high-scale applications.

⚠️ Development Focus

We are currently prioritizing: * Developer Experience (DX): Simplifying the integration of persistent memory into existing agentic frameworks. * Performance: Minimizing latency in hybrid retrieval pipelines. * Accuracy: Reducing hallucinations by providing higher-quality, structured context.

Last Updated: January 2026