Last updated: November 3, 2026
Hindsight Models are three new models from Vectorize for agent memory: hs-decide-1-flash answers typed yes, which-one and how-likely questions, hs-rerank-1 ranks memories and hs-embed-1 creates embeddings. They ship as open weights and through a hosted API priced from $0.005 per million input tokens.
Contents
- What’s new at a glance
- What a decision model does
- API pricing and endpoints
- How to turn it on in Hindsight 1.2
- Using the models in your own app
- Published results and trade-offs
- How it compares
- Limitations
- Frequently asked questions
Vectorize announced the models on November 3, 2026, in the Introducing Hindsight Models post, alongside Hindsight 1.2.0, the release of its open-source memory server that adds support for them. The Hindsight GitHub repository had about 58,400 stars when we checked.
Much of what a memory system asks an LLM is judging, not writing: is this fact new, is this evidence enough? A small decision model answers those faster and cheaper.
What’s new at a glance
| Model | What it does | Size and base | License | Context |
|---|---|---|---|---|
hs-decide-1-flash |
Typed judgments: a choice, a score or how likely | 0.8B, built on Qwen3.5-0.8B | Apache 2.0 | 65,536 tokens |
hs-rerank-1 |
Orders documents by relevance to a query | 114M, built on mmBERT-small | MIT | 8,192 tokens |
hs-embed-1 |
Turns text into 384-dimension vectors | 112M, built on bekko-embedding-v1-a25m | MIT | 8,192 tokens |
All three are published in the hindsight-ai organization on Hugging Face and, according to Vectorize, work in English, Spanish, French, German, Portuguese, Japanese, Korean and Chinese. The reranker and embedding model ship an int8 ONNX export and are fast on CPU.
What a decision model does
The Decision Models documentation describes a model that answers short, typed questions in well under a second and returns a choice or a probability instead of prose, so there is nothing to parse. Hindsight now asks it at four points:
- Recall: it ranks each candidate memory by whether it would help answer the question, and can drop unhelpful ones (prune) or repeats (dedup).
- Consolidation: it settles facts that only restate an existing observation or are chatter, so only real changes reach the LLM.
- Reflect: it trims the evidence and judges whether it already answers the request before another search round runs.
- Knowledge pages: it asks whether each new memory could change a page, and skips the LLM when none could.
Vectorize reports a median of 167 ms per decision-model request (350 ms at p90), compared with 1 to 7 seconds for the LLM call it replaces. The model is optional. If it fails, consolidation, reflect and page refreshes fall back to the LLM path; recall is the exception, because the decision model is its ranker when enabled.
API pricing and endpoints
Hindsight Inference is part of Hindsight Cloud and shares the same organization, teams and credit balance. None of the models generates text, so you pay for input tokens only.
| API | Endpoint | Model | Drop-in for | Price per million input tokens |
|---|---|---|---|---|
| Decisions | POST /v1/systemone |
hs-decide-1-flash |
System One | $0.02 |
| Reranker | POST /v1/rerank |
hs-rerank-1 or hs-decide-1-flash |
Cohere Rerank | $0.015 / $0.02 |
| Embeddings | POST /v1/embeddings |
hs-embed-1 |
OpenAI Embeddings | $0.005 |
New accounts start with $5 of free credit. Vectorize says that covers the decision model’s part of about 4,000 reflect operations on its evals.
How to turn it on in Hindsight 1.2
In open-source Hindsight 1.2 the decision model is off until you configure it. The hosted path needs no GPU.
- Upgrade the server to Hindsight 1.2.0. Vectorize recommends that everyone on 1.1.x upgrade.
- Run
hindsight-admin cloud loginto sign in to Hindsight Cloud and save a key. Alternatively, setHINDSIGHT_API_DECISION_MODEL_API_KEYwith a key from Hindsight Cloud. - Set
HINDSIGHT_API_DECISION_MODEL_PROVIDER=hindsight. This turns on the decision model for consolidation, reflect and knowledge pages. - Optional: set
HINDSIGHT_API_RECALL_DECISION_MODEL_PROVIDER=hindsightto rank recall with it, and setHINDSIGHT_API_RECALL_DECISION_MODEL_PRUNE=trueandHINDSIGHT_API_RECALL_DECISION_MODEL_DEDUP=trueto drop unhelpful and repeated memories. - Optional: set
HINDSIGHT_API_RERANKER_PROVIDERandHINDSIGHT_API_EMBEDDINGS_PROVIDERtohindsightfor hosted models, or tolocalwithhindsight-ai/hs-rerank-1andhindsight-ai/hs-embed-1to run them inside the API process.
Prune and dedup are off by default. You can also turn them on per request with "rerank": {"prune": true, "dedup": true}. Note that switching embedding models means re-embedding a bank’s stored vectors.
Two 1.2.0 changes matter even if you skip the decision model. The local and ONNX rerankers now default to hs-rerank-1 instead of cross-encoder/ms-marco-MiniLM-L-6-v2; set HINDSIGHT_API_RERANKER_LOCAL_MODEL to keep the old one. And tool calls now go through a response schema, so an OpenAI-compatible server must support structured output. If you self-host models, our vLLM 0.32 guide covers recent serving changes.
Using the models in your own app
You do not need the memory server. Sign in at inference.hindsight.vectorize.io, try each model in the Playground, which shows the request as curl, Python, Cohere SDK or OpenAI SDK code, and create a key under API Keys. Point a Cohere client at https://inference.hindsight.vectorize.io or an OpenAI client at https://inference.hindsight.vectorize.io/v1 and change the model name.
A useful pattern: if you already call Cohere Rerank, set the model to hs-decide-1-flash and the endpoint returns only relevant documents rather than ranking all of them. Our Cohere Rerank 5 guide explains the hosted reranker it replaces. For “is this new or already known” checks in your own app, a single Decisions API call can stand in for an LLM prompt.
Published results and trade-offs
Vectorize’s numbers come from its own system evals on Hindsight 1.2.0, graded by gemini-3.1-flash-lite. They are vendor results, not independent benchmarks.
- Quality rose for six of seven LLMs tested across Hindsight, and cost fell for five. On gemini-3.8-flash, the operations the decision model takes part in cost 2.5x less and scored 2.2 points higher.
- glm-5.3-flash with the decision model scored 87.1 at $7.72 per thousand cases, against 89.0 at $147 for gemini-3.8-flash alone.
- Reflect gives the biggest saving (4.6x on gemini-3.8-flash) but quality dipped about 3 points for four LLMs and fell from 94.1 to 85.3 for gemini-3.1-flash-lite.
- The two cheapest LLMs cost more with the decision model, because its own tokens outweigh what it saves, though their answers improved sharply.
How it compares
Against other open decision models of its size, Vectorize reports the best mean over eight benchmarks (75.6, against 74.5, 73.4 and 66.6), with its lead in memory judgments. It trails on general decision benchmarks, such as the public Decision Index (36.1 against 53.8 for StartLux). For a general-purpose hosted alternative, see our Perplexity Decisions API guide.
If your memory layer runs on a hosted model such as Claude Haiku 5.5 or a low-cost one such as DeepSeek V4.1 Flash, test both setups on your own data first, especially for reflect-heavy workloads.
Limitations
- Reflect and knowledge-page refresh currently need the hindsight or typesafe provider for the decision model; other providers keep the LLM-only path.
- With the decision model as recall ranker, a recall fails if the model cannot be reached after retries.
- Decision-model ranking replaces recall’s usual recency, temporal and proof-count boosts.
- The 1.2.0 release removes several settings, including HINDSIGHT_API_LLM_SUPPORTS_MAX_ITEMS, so check your configuration before upgrading.
Frequently asked questions
What are Hindsight Models?
Hindsight Models are three small models from Vectorize, released November 3, 2026, for agent memory: hs-decide-1-flash, a 0.8B decision model; hs-rerank-1, a 114M reranker; and hs-embed-1, a 112M embedding model. They are hosted through Hindsight Inference and published as open weights.
How much does Hindsight Inference cost?
Vectorize lists input-token pricing only, because none of the models generates text: $0.02 per million tokens for decisions, $0.015 per million for reranking with hs-rerank-1 ($0.02 with hs-decide-1-flash) and $0.005 per million for embeddings. New accounts get $5 of free credit.
Are the Hindsight Models open source?
The weights are on Hugging Face in the hindsight-ai organization. The decision model is under Apache 2.0, and the reranker and embedding model are under MIT. The Hindsight memory server itself is MIT-licensed on GitHub.
Do I need a GPU to use them?
Not for the hosted path: one login and one environment variable turn on the decision model with no GPU. The reranker and embedding model also run on a CPU. Serving the decision model yourself uses vLLM.
Can I use Hindsight Inference without the Hindsight memory server?
Yes. The Decisions, Reranker and Embeddings endpoints work as drop-ins for System One, Cohere Rerank and OpenAI Embeddings clients. You point the client at the Hindsight Inference base URL and change the model name.
Sources
- Introducing Hindsight Models (Hindsight blog)
- What’s new in Hindsight 1.2.0 (Hindsight blog)
- Decision Models (Hindsight documentation)
- hindsight-ai models (Hugging Face)
- vectorize-io/hindsight (GitHub)
Image: Hindsight logo from the vectorize-io/hindsight repository, MIT License, placed on a white background and resized. Hindsight is a product of Vectorize, Inc. FreeAnswers is not affiliated with Vectorize.