SMELT slashes agent token waste 95%

SMELT targets the "silent killer" of agentic workflows: context bloat from re-sending static data in every turn. By treating workspace files as a schema-aware database rather than raw text, it massively reduces latency and costs for both local and hosted models.

• –Achieves a 95% token reduction (e.g., 1,373 down to 73 tokens), directly improving model focus and inference speed.
• –Avoids the 30% overhead typical of naive JSON conversion through a specialized layered compression approach.
• –Benchmarked on Qwen 3.5 122B, proving viability for optimizing high-parameter local MoE models on consumer hardware like the M3 Ultra.
• –Specifically addresses design flaws in frameworks like OpenClaw that lack efficient workspace memory management.
• –The project includes a formal research paper on Zenodo, providing technical depth beyond standard Reddit "show and tell" posts.