HindsightFrom the builder
Introducing Hindsight Models: memory 2.5x cheaper and 2x faster, with better answers
Hindsight released three models for agent memory, led by the decision model hs-decide-1-flash that takes over small judgments in consolidation, reflect, and knowledge pages so the LLM can focus on writing. Across seven LLMs, six scored better and five got cheaper, with clear cost and latency drops on gemini-3.8-flash for the operations the decision model joins.
The company also launched the Hindsight Inference API and open weights, billed on input tokens with free credit for new accounts, Apache for the decision model and MIT for the reranker and embedder. Companion models hs-rerank-1 and hs-embed-1 target memory retrieval, run on CPU, and cover eight languages.
The LLM writes, the decision model judges. hs-decide-1-flash takes over the small judgments inside Hindsight: better answers from six of seven LLMs, cheaper for five.
Saved . A shortened excerpt from the official blog; the original publisher retains authorship.