Skip to main content

What's new in Hindsight 1.2.2

Browse a bank's observations as a folder of markdown files, a webhook when a mental model refreshes, newer default LLM models, and a long round of reliability fixes

By Nicolò Boschi
· 5 min read

Hindsight 1.2.2 lets you browse everything a bank has learned as a folder of markdown files, sends a webhook when a mental model finishes refreshing, and moves the default LLM models to newer, cheaper ones. Most of the release is reliability work: retain keeps tags it used to drop, reflect handles more model replies, and Gemma 4 models stop losing facts. Everyone on 1.2.1 should upgrade — see the changelog for the full list.

Observation Tree​

The Observation Tree shows every observation in a bank as a folder of markdown files: one file per topic, and an index.md in each folder that says what is inside. An agent can ls and grep it to see what the bank knows before it has a question to ask. A person can browse it to check what was learned.

Topics are grouped by the entities the observations mention, with no LLM call, and the same bank always gives the same paths. The files follow the Open Knowledge Format.

In the control plane, open a bank and go to Pages → Observations to browse it, search it, or download it as a .zip. Over the API, call GET /v1/default/banks/{bank_id}/observations/tree, with tags, tags_match or tag_groups to build it from matching observations only. It is experimental: the endpoint and file layout may still change.

Webhook when a mental model refreshes​

A new mental_model.refreshed webhook event fires when a mental-model refresh finishes. Use it to push the updated page somewhere else, or to clear a cache, instead of polling.

Newer default models​

If you don't set HINDSIGHT_API_LLM_MODEL, the default model for your provider changed to a newer, cheaper one:

ProviderBeforeNow
OpenAI, LiteLLM, Requesty, fallbackgpt-4o-minigpt-5-nano
Anthropicclaude-haiku-4-5claude-haiku-5-5
Claude Codeclaude-sonnet-4-5-20250929claude-sonnet-5-5
Geminigemini-3.5-flashgemini-3.8-flash

Set HINDSIGHT_API_LLM_MODEL to keep your current model.

Smaller additions​

  • Extra-body settings merge: an object in HINDSIGHT_API_LLM_EXTRA_BODY now adds to the one Hindsight sends instead of replacing it, so you can set e.g. Claude's effort or OpenAI's verbosity without losing the response schema.
  • Cheaper reflect on Claude: Claude's prompt cache now carries across a reflect's turns, so long reflects resend less.
  • Faster local GPU inference: local models no longer run a garbage collection on every call.
  • Provider refusals return 422 instead of 500.

Reliability​

  • An append without tags keeps the document's tags; reordering tags no longer counts as a change.
  • Changing a memory's tags now updates the observations built from it first, so they don't keep the old tags.
  • Decision-model ranking in reflect uses the window you configured.
  • Heavy CPU work no longer blocks the API while it runs, so other requests keep answering.
  • Re-retaining a document without observation_scopes keeps the bank's default scope on the facts it keeps.
  • Cloning a large bank no longer runs as one long transaction, so it doesn't time out.
  • File downloads work for bank ids with non-ASCII characters, and the control plane opens a bank whose id ends like a file name (mybank.txt).
  • hindsight retain-files keys each document on its full path, so two files with the same name no longer overwrite each other.
  • Gemma 4 on vLLM keeps retain's facts, a model that echoes the schema back is retried on the fallback model, and consolidation stops the misses Gemma 4 made on 1.2.1.
  • Retain retries its completion write when the database is briefly overloaded; the worker no longer spends a retry on that, and stops deferring after 180 tries.
  • Reflect reads a native tool call as a reply, re-asks once on an empty answer, and HINDSIGHT_API_REFLECT_MAX_COMPLETION_TOKENS now applies to every turn.
  • Reflect places a new section after another new one correctly.
  • A reranker or embedder blip during reflect no longer pauses a mental model's automatic refresh, and refresh stops running to the token cap.
  • MCP retain rejects empty content.
  • hindsight-admin reembed tells a leftover column apart by model, not only by size.
  • HINDSIGHT_API_RERANKER_LOCAL_FP16 now works on CPU.
  • On Windows, the server starts when its output is redirected.
  • On Oracle, observation scope counts and exact tag matching work.
  • Tracing records a span for timed-out retries and cancelled LLM calls.
  • The control plane's reflect trace shows steps in the order they ran.
  • Dependency bumps for new CVEs (dnspython, soupsieve, Next.js).

Database migrations: three migrations run on startup. They only create and change the knowledge-bank tables, which stay unused unless knowledge banks are turned on.