Skip to main content

What's new in Hindsight 0.8.6

· 5 min read
Nicolò Boschi
Nicolò Boschi
Hindsight Team

Hindsight 0.8.6 builds on 0.8.5 with a focus on control and visibility: browse a bank by entity and tag, and tune how hard the model thinks per operation.

Browse Memory by Entity and Tag​

Memory is only as useful as your ability to inspect it. This release makes browsing a bank far more direct:

  • Entity timeline. Pick an entity — a person, a company, a project — and see everything Hindsight knows about it laid out in order, from the first mention onward. It's the fastest way to check what your agent actually learned about someone, and when.
  • Filter memories by linked entity. The timeline is backed by an API filter, so the same "everything about this entity" view is available to your own tools, not just the control plane.
  • API version in the sidebar. The control plane shows which API version it is talking to, so you always know what you're pointed at.

A Documents Table Built for Scanning​

The documents list is where you go to answer "what has this bank actually ingested?" — and until now it only let you search by document ID.

  • Filter by tag. A tag filter with autocomplete and an any/all toggle, so you can narrow a bank down to the documents you care about instead of scrolling.
  • A table organized around what you scan for. The last-updated time now sits under the document ID, tags and metadata share a column, and size and memory-unit counts get their own aligned columns. Long document IDs no longer push everything else off the row.
  • A filter you can always clear. The filter toolbar used to disappear when a filter matched nothing, which left you stuck with an empty list and no way out.
  • Consistent chips for tags, entities, and metadata. These three used to be styled differently in nearly every view — tags were blue in one place and amber in another, and entities were indistinguishable from tags. They now render the same way everywhere, told apart by shape rather than by color, with the accent color reserved for a filter that's actually active.

Per-Operation Model Control​

Reasoning models are not equally useful at every stage of a memory pipeline. Extraction benefits from care; a short reflect answer usually does not, and paying for reasoning tokens there is pure latency and cost.

0.8.6 lets you set the reasoning effort per operation instead of once for the whole deployment, with HINDSIGHT_API_RETAIN_LLM_REASONING_EFFORT, HINDSIGHT_API_REFLECT_LLM_REASONING_EFFORT, and HINDSIGHT_API_CONSOLIDATION_LLM_REASONING_EFFORT. Anything you don't set keeps following the global HINDSIGHT_API_LLM_REASONING_EFFORT, so existing deployments behave exactly as before.

More of the pipeline became tunable alongside it:

  • Configurable embedding thresholds, so you can decide how similar is similar enough for your data.
  • store_document_text per bank, so you can keep raw document text out of specific banks while keeping it everywhere else.

Operate with More Signal​

  • Health checks that explain themselves. When /health stalls, Hindsight now reports whether the event loop is blocked or the database connection pool is exhausted — the difference between a restart that helps and a restart that thrashes.
  • Worker slot settings say what they do. Per-type slot settings are now HINDSIGHT_API_WORKER_<TYPE>_RESERVED_SLOTS: they reserve a guaranteed minimum for an operation type, they never cap it. The old *_MAX_SLOTS names keep working and log a deprecation warning, so nothing changes until you rename them.

Build on Hindsight​

  • Extension-owned tables. Extensions can declare bank-scoped tables that Hindsight provisions automatically for every bank.
  • Custom headers in the TypeScript client, for gateways and proxies that need them.
  • Russian temporal period rules, for parsing dates and periods in Russian-language content.
  • Python 3.14 is supported.

0.8.6 also carries a large round of reliability fixes across workers, graph maintenance, curation, and platform support (Oracle, Bedrock, Kubernetes, Apple Silicon). Self-managed deployments should upgrade — several of those fixes protect background work from stalling or duplicating. See the changelog for the full list.