v0.4.17 v0.4.18

Name: hindsight
Author: vectorize-io

low risk

vectorize-io/hindsightMar 10, 2026 → Mar 13, 20263 days between releases

What's New

Adds MiniMax provider and performance improvements using UNION ALL

New Features

MiniMax LLM provider support
Jina MLX reranker provider for Apple Silicon
Compound tag filtering via tag_groups
hindsight-api-slim packages
Configurable recall max query tokens via environment variable

Improvements

Performance optimization replacing window-function retrieval with UNION ALL
Per-bank HNSW indexes

Bug Fixes

Cancel in-flight async ops when bank is deleted
Register embedded profiles in CLI metadata on daemon start
Keep recalled memories out of the visible transcript

Upgrade Notes

You can now configure recall max query tokens via the RECALL_MAX_QUERY_TOKENS environment variable. New slim packages (hindsight-api-slim and hindsight-all-slim) are available. Ensure your configuration supports the new tag_groups filtering syntax if you use tag-based queries.

The Two Releases

v0.4.17

Mar 10, 2026· 8 months ago

Introduces history tracking for mental models and observations, improves file parsing with fallback chains, and adds manual retry capabilities for async operations.

Full release notes

v0.4.18

Mar 13, 2026· 7 months ago

Focuses on performance improvements using UNION ALL queries, adds new LLM providers like MiniMax, and introduces reranker support for Apple Silicon.

Full release notes