Support for the Hindsight models, a new default reranker, prune and dedup controls for decision-model recall, images on knowledge pages, a better LLM request view, and a long round of reliability fixes
Hindsight 1.2.0 adds support for the Hindsight models, makes hs-rerank-1 the default reranker, gives recall finer control over what the decision model keeps, shows the images behind knowledge pages, and makes the LLM request view easier to read. Tool calls now go through a response schema on every provider, which makes reflect far more reliable on self-hosted models. Most of the release is reliability work across retain, consolidation, reflect, refresh and the worker. Everyone on 1.1.x should upgrade — see the changelog for the full list.
- [**Hindsight models**](#hindsight-models) : A decision model, a reranker and an embedding model built for agent memory;`hs-rerank-1` is the new default reranker.
- [**Decision-model recall: prune and dedup**](#decision-model-recall-prune-and-dedup) : Drop irrelevant and duplicate memories, per request or per bank.
- [**Tool calls through a response schema**](#tool-calls-through-a-response-schema) : Reflect works on servers that ignore forced tool choice.
- [**Images on knowledge pages**](#images-on-knowledge-pages) : See the images behind a page's evidence.
- [**A better LLM request view**](#a-better-llm-request-view) : Retention, an hourly chart, and a split by model or operation.
- [**More settings**](#more-settings) : Vision call settings, embedding extra body, a default time zone.
- [**Helm**](#helm) : Optional API command, startup probe, strategy and token mounting.
Hindsight models #
1.2.0 supports the Hindsight models: hs-decide-1-flash, a decision model that takes over the small judgments in consolidation, reflect and knowledge pages, plus hs-rerank-1 and hs-embed-1. They're off until you set them up: run hindsight-admin cloud login and set HINDSIGHT_API_DECISION_MODEL_PROVIDER=hindsight. See Decision Models for every setting.
The default reranker changes. The local and ONNX rerankers now use hindsight-ai/hs-rerank-1, multilingual (eight languages) and trained for Hindsight recall, instead of cross-encoder/ms-marco-MiniLM-L-6-v2. It downloads on first start. To keep the old model on the local provider, set HINDSIGHT_API_RERANKER_LOCAL_MODEL=cross-encoder/ms-marco-MiniLM-L-6-v2.
The typesafe reranker provider is deprecated in favour of the shared decision model.
Decision-model recall: prune and dedup #
When the decision model reranks recall, it can now also drop memories that don't help answer the query (prune) and memories that repeat one already returned ( dedup). Turn them on per request (HTTP and MCP recall), or per bank with HINDSIGHT_API_RECALL_DECISION_MODEL_PRUNE and HINDSIGHT_API_RECALL_DECISION_MODEL_DEDUP. The decision model now also sees each memory's date, so time-sensitive questions rank better.
Tool calls through a response schema #
Reflect's tool calls now go through a response schema instead of the provider's native tool calling. Some OpenAI-compatible servers (for example vLLM with some open models) accept tools but don't honour a forced tool choice, so reflect's forced steps were often skipped. Those servers do honour a response schema, so reflect now behaves the same on every provider.
Images on knowledge pages #
When a knowledge page's evidence comes from retained images, the page now shows them next to that evidence in the control plane.
A better LLM request view #
The LLM request view says how many days of history are kept, adds an hourly chart for the last 24 hours, and splits usage by model or by operation. It now also records embedding, rerank and decision-model calls, not only LLM calls. Set HINDSIGHT_API_LLM_TRACE_DELETE_WITH_DOCUMENT=true to delete a document's request history when the document is deleted; deleting a bank always deletes its history.
More settings #
- Vision calls get their own reasoning effort, timeout, retries and extra body (
HINDSIGHT_API_VLM_*). - OpenAI-compatible embeddings can send an extra request body (
HINDSIGHT_API_EMBEDDINGS_OPENAI_EXTRA_BODY). - Default time zone (
HINDSIGHT_API_DEFAULT_TIMEZONE): retain timestamps without a UTC offset are read in this zone.
Helm #
The chart adds an optional API command, a startupProbe, a deployment strategy, and control over automountServiceAccountToken. Docker and Helm give the API 60 seconds to finish in-flight work on shutdown.
Breaking changes: no provider is sent native tools any more, so an OpenAI-compatible server must support a response schema (structured output). HINDSIGHT_API_LLM_SUPPORTS_MAX_ITEMS, HINDSIGHT_API_LLM_STRUCTURED_OUTPUT_FORCED_TOOL and HINDSIGHT_API_REFLECT_PROMPT_CACHE_ENABLED are removed; remove them from your configuration. On the LiteLLM-backed providers (litellm, litellmrouter, bedrock), set HINDSIGHT_API_LLM_STRUCTURED_OUTPUT_IN_PROMPT=true to give the schema in the prompt instead.