Skip to main content
Version: v0.1 (Stable)

Hybrid Semantic Memory & Local Vector RAG

ActonOS implements a zero-cloud-dependency Hybrid Memory Engine that combines exact lexical search, dense vector embeddings, and cognitive forgetting curves.

1. The Local ONNX Embedding Helper (embeddingd)​

To keep the main actond daemon a 100% pure static Go binary (CGO_ENABLED=0), local ONNX neural inference runs in a separately packaged loopback helper daemon:

  • Model: intfloat/multilingual-e5-small (384-dimensional vector embeddings).
  • Languages: 100+ languages including English, Vietnamese, Chinese, Japanese, French, Spanish, German.
  • Latency: Under 15ms per chunk on low-power Intel N100 MiniPCs.

2. Durable Debounce Queue​

When an agent or user writes multiple files rapidly, ActonOS queues paths in a persistent SQLite table (workspace_embedding_queue):

  • Debounce window: 60 seconds.
  • Eliminates redundant vector calculations during rapid compilation or file editing.
  • Survives unexpected host crashes without losing indexing state.

3. Cognitive Decay (Ebbinghaus Forgetting Curve)​

To prevent irrelevant historical logs from crowding LLM context windows, memories are weighted using the Ebbinghaus exponential decay model:

R(t) = exp(-t / S)

Where:

  • R(t): Memory retention strength (0.0 to 1.0).
  • t: Elapsed time since last retrieval or reinforcement.
  • S: Memory stability score (boosted whenever an agent references the memory item).