Sillage Sillage
Sillage is a gradient-free memory tool for frozen language models that provides persistent learning in a fixed-size memory footprint. It utilizes Hebbian matrices, a semantic tier, and rank-16 adapters to allow models to learn from new text without fine-tuning or growing index stores.
Version 1.9.2 released in late August 2026, consolidating previous research scripts into a unified CLI and refining memory readout calibration protocols.
As of
- Fixed-size memory footprint that remains constant regardless of data volume
- No fine-tuning or gradients required for learning
- Compatible with any causal language model
- CPU-first execution with support for GPU acceleration
- Includes local full-text search and session journaling
Users who need a model to remember and learn from private technical documents or repetitive text on a single machine without an unbounded vector database.
Applications involving long, low-repetition narrative text where traditional RAG or kNN-LM approaches typically outperform surface-level recurrence memory.
MIT licensed; free to use as a command-line tool or library.
An innovative, research-backed solution for adding efficient, fixed-memory persistence to frozen models, ideal for developers working with repetitive technical corpora on local hardware.
Is Sillage free?
Yes - Sillage is Open Source. MIT licensed; free to use as a command-line tool or library.
What is Sillage best for?
Users who need a model to remember and learn from private technical documents or repetitive text on a single machine without an unbounded vector database.
Who makes Sillage?
Sillage is developed by Abderrahmane Sghairi. It is listed in the Terminal & CLI Agents category on ai.dosa.dev.
Favorite this tool to revisit it later, or Zap it to contribute to the public vote count.
Content on this page is AI-generated. Please verify details with the vendor's website for accuracy.