Vectorize releases Hindsight models: an open 0.8B decision model, a reranker and an embedder for agent memory

Vectorize, which makes the Hindsight agent memory system, has released three small models for it. The main one, hs-decide-1-flash, is a decision model: instead of generating text, it answers typed questions (pick one option, rate on a scale, yes or no) with a probability for each answer. Vectorize’s launch post on 3 November frames it as “the LLM writes, the decision model judges”: a memory system asks many small questions, such as whether a fact is new or whether retrieved evidence is enough, and those no longer need a full LLM call.

The three models

Model	Job	Size	Built on	License
hs-decide-1-flash	Typed judgments	0.8B	Qwen3.5-0.8B	Apache 2.0
hs-rerank-1	Reranking	114M	mmBERT-small	MIT
hs-embed-1	384-dimension embeddings	112M	bekko-embedding-v1	MIT

All three are open weights in the hindsight-ai organization on Hugging Face and cover eight languages, including English, Japanese, Korean and Chinese. The decision model card lists a 65,536-token context and ships a vLLM plugin that adds a /v1/systemone endpoint. Vectorize says the reranker and embedder are fast enough for a CPU. The same models are also hosted on a paid Hindsight Inference API.

What changes in Hindsight

Hindsight 1.2.0, the open-source MIT-licensed release that shipped the same day, can use the decision model in consolidation, reflect and knowledge-page refreshes, and in recall if you enable it. It is opt-in: the docs say that with no provider set (the default), every operation keeps its LLM-only path. Besides Vectorize’s own hosted model, the docs list Jev, Cloudflare’s Clef and OpenAI Decisions as providers, though reflect and knowledge pages support only Vectorize’s and Jev for now. One default does change: 1.2.0 makes hs-rerank-1 the default reranker.

The results, as Vectorize reports them

These come from Vectorize’s own system evals, run with seven LLMs alone and with the decision model, all graded by one LLM judge (gemini-3.1-flash-lite, which is also one of the seven, so its own rows are self-graded):

• Across Hindsight: quality goes up for six of the seven LLMs, and five get cheaper.
• The headline: on gemini-3.8-flash, counting only the operations the decision model takes part in, cost falls 2.5 times, time in model calls falls 1.9 times and quality rises 2.2 points.
• Cheap LLMs gain most: glm-5.3-flash goes from 73.8 to 87.1, about 2 points below gemini-3.8-flash alone at 15 to 19 times less cost. It and deepseek-v4.1-flash, the two cheapest, cost more with the decision model than without, because their own calls are so cheap.
• Speed: a median of 167 ms per decision-model request (350 ms at the 90th percentile), against 1 to 7 seconds for an LLM call.
• Trade-offs: in reflect, quality dips about 3 points for four of the LLMs and drops from 94.1 to 85.3 for gemini-3.1-flash-lite.

Why this matters

If you run Hindsight yourself, upgrading to 1.2.0 changes your reranker but leaves the decision model off, so you can try it per operation and keep the LLM path where reflect quality matters most. Where Tencent’s Hy Memory 2.0 pairs its memory system with a 30B model, Hindsight’s helper is 0.8B and can be swapped for a third-party decision model such as Cloudflare’s Clef. Re-run the comparison on your own memory workload before trusting the 2.5x figure, since it comes from Vectorize’s evals and its best LLM pairing.