vectorize-io/hindsight v1.2.0 on GitHub

What's Changed

docs: changelog and blog post for v1.1.0 by @nicoloboschi in #5935
docs(blog): ChatGPT sign-in recording in the v1.1.0 post by @nicoloboschi in #5939
fix(parsers): strip UTF-8 byte-order mark in markitdown text files by @nicoloboschi in #5949
fix(embeddings): don't fail the call when the provider omits prompt_tokens by @nicoloboschi in #5951
fix(reflect): treat a non-list id field on done as empty instead of failing the reflect by @nicoloboschi in #5952
fix(reflect): treat a non-list memory_ids on expand as a tool error instead of failing the reflect by @nicoloboschi in #5954
fix(banks): delete a bank's LLM request history with the bank by @nicoloboschi in #5948
fix(metrics): a hard client disconnect no longer counts as a failed operation by @nicoloboschi in #5955
fix(embeddings): OpenAI-compatible embeddings send the hindsight User-Agent by @nicoloboschi in #5950
fix(admin): repair-bank removes observations whose source facts are gone (#5937) by @nicoloboschi in #5957
fix(docker): build the control plane from the root lockfile by @nicoloboschi in #5959
fix(banks): delete a bank's maintenance queue rows with it; test that a bank delete leaves no row in any table by @nicoloboschi in #5958
test: fix two tests that fail outside the default setup by @nicoloboschi in #5962
fix(audit): always log destructive operations and name the tool in MCP call logs by @nicoloboschi in #5961
feat(openclaw): let cron, heartbeat and subagent sessions use memory when opted in (#5915) by @nicoloboschi in #5963
feat(openclaw): offer Hermes's Hindsight tools on the agent's own bank (#5635) by @nicoloboschi in #5965
fix(obsidian): unstick sync on already-deleted documents; option to date facts by last edit by @nicoloboschi in #5967
fix(obsidian): bump manifest.json + versions.json on release by @nicoloboschi in #5972
fix(consolidation): an older fact no longer overwrites a newer same-day amendment (#5938) by @nicoloboschi in #5966
feat(extensions): ModelCallExtension, every model call with its input and output by @nicoloboschi in #5936
test(system): model call extension story allows a recall with no rerank call by @nicoloboschi in #5978
fix(docker): control-plane build no longer flakes on fonts or a missing SWC binding by @nicoloboschi in #5987
feat/helm api command probe strategy by @nicoloboschi in #5988
docs(integrations): add Hindsight Ingestion community integration by @nicoloboschi in #5992
fix(tests): run migration tests on the shared test server by @nicoloboschi in #5989
fix(openclaw): save a run from where it started, not from its last user message (#5994) by @nicoloboschi in #5995
fix(consolidation): rebuild the VectorChord keyword index on observation rewrites and folds (#5299) by @nicoloboschi in #5991
fix(consolidation): a skipped retry ends cancelled with superseded_by, not failed (#5434) by @nicoloboschi in #5990
fix(control-plane): restore files a stray prettier run reformatted in #5989 by @nicoloboschi in #6002
fix(hermes): six small plugin bugs by @nicoloboschi in #6000
fix(coding-agents): load on OpenCode 2.1 by package name, unquoted Antigravity commands on Windows by @nicoloboschi in #6003
docs: cheap memory model with failover recipe by @nicoloboschi in #5997
feat(openclaw): add Hindsight's knowledge-base tools to the agent tools (#5635) by @nicoloboschi in #6004
fix(api): don't load fastmcp when MCP is disabled by @nicoloboschi in #6001
chore(integrations): remove the per-agent coding packages Coding Agents replaced by @nicoloboschi in #6008
fix(hermes): don't lose automatic retains when a write fails by @nicoloboschi in #6013
ci(hermes): install Hermes and the plugin the way users do by @nicoloboschi in #6014
refactor(engine): stop the engine importing the API layer and untangle import cycles by @nicoloboschi in #5993
fix(consolidation): stop sending each source memory's context (#5996) by @nicoloboschi in #6018
fix(reflect): the search-query rewrite never ends reflect with an empty answer by @nicoloboschi in #6022
docs: remove versioned docs for 0.7 and 0.8 by @nicoloboschi in #6031
fix(consolidation): drop create/update actions with empty text by @nicoloboschi in #6024
chore(deps): bump toml and vitest in coding-agents lockfile by @cdbartholomew in #6027
chore(deps): bump vitest in ai-sdk lockfile by @cdbartholomew in #6028
chore(deps): bump vitest in chat lockfile by @cdbartholomew in #6029
chore(deps): bump vitest in cloudflare-oauth-proxy lockfile by @cdbartholomew in #6030
chore(deps): bump vitest in eliza lockfile by @cdbartholomew in #6032
chore(deps): bump vitest in eve lockfile by @cdbartholomew in #6033
chore(deps): bump vitest in flowise lockfile by @cdbartholomew in #6034
chore(deps): bump vitest in n8n lockfile by @cdbartholomew in #6035
chore(deps): bump vitest in obsidian lockfile by @cdbartholomew in #6037
chore(deps): bump vitest in openclaw lockfile by @cdbartholomew in #6038
chore(deps): bump opentelemetry-api in agent-framework lockfile by @cdbartholomew in #6041
chore(deps): bump opentelemetry-api in autogen lockfile by @cdbartholomew in #6042
chore(deps): bump opentelemetry-api in crewai lockfile by @cdbartholomew in #6043
chore(deps): bump opentelemetry-api in pydantic-ai lockfile by @cdbartholomew in #6044
chore(deps): bump pypdf and opentelemetry-api in superagent lockfile by @cdbartholomew in #6046
chore(deps): bump vitest in nemoclaw lockfile by @cdbartholomew in #6036
chore(deps): bump opentelemetry-api in strands lockfile by @cdbartholomew in #6045
chore(deps): bump vitest in paperclip lockfile by @cdbartholomew in #6039
chore(deps): bump next, dompurify, vitest in root lockfile by @cdbartholomew in #6026
chore(deps): bump opentelemetry packages in root uv.lock by @cdbartholomew in #6040
fix(worker): shutdown releases only the schemas the worker claimed in by @cdbartholomew in #6047
fix(reflect): return whole knowledge pages up to a token budget, not one page under 4k by @nicoloboschi in #6009
feat(extensions): operation runs - each operation's wall time and where it went by @nicoloboschi in #6048
fix(coding-agents): small retain, recall and survey fixes by @nicoloboschi in #6050
fix(coding-agents): put the memory block in the user message, not the system prompt by @nicoloboschi in #6053
fix(coding-agents): a session saves each turn once, across bank moves and forks by @nicoloboschi in #6052
fix(control-plane): replace removed next lint with eslint and Next's rule set by @nicoloboschi in #6049
fix(coding-agents): launch the codebase survey on Windows and stop retrying forever by @nicoloboschi in #6054
chore(deps): bump opentelemetry-api in google-adk lockfile by @cdbartholomew in #6066
chore(deps): bump zod, body-parser, undici, fast-uri, serialize-javascript in root lockfile by @cdbartholomew in #6079
chore(deps): bump undici in cloudflare-oauth-proxy lockfile by @cdbartholomew in #6080
chore(deps): bump undici and fast-uri in coding-agents lockfile by @cdbartholomew in #6081
chore(deps): bump undici in eve lockfile by @cdbartholomew in #6082
chore(deps): bump zod to v4 in flowise integration by @cdbartholomew in #6083
chore(deps): bump @paperclipai/plugin-sdk (zod v4) in paperclip lockfile by @cdbartholomew in #6084
chore(deps): bump undici and serialize-javascript in zapier lockfile by @cdbartholomew in #6085
chore(deps): bump filelock in crewai lockfile by @cdbartholomew in #6086
chore(deps): bump filelock in litellm lockfile by @cdbartholomew in #6087
chore(deps): bump filelock in smolagents lockfile by @cdbartholomew in #6088
chore(deps): bump setuptools in llamaindex lockfile by @cdbartholomew in #6090
chore(deps): bump strands-agents in strands lockfile by @cdbartholomew in #6089
fix(retain): don't checkpoint documents a buffering retain session hasn't committed (#6099) by @nicoloboschi in #6100
ci: swap retired gpt-4.1-nano for gpt-6-luna in LLM acceptance matrix by @nicoloboschi in #6102
fix(control-plane): hide the next-server console window on Windows by @nicoloboschi in #6104
fix(models): ONNX loaders no longer crash-loop startup next to a local model build by @nicoloboschi in #6105
fix(bank-template): export keeps entity_labels stored in the attributes envelope by @nicoloboschi in #6106
fix(llm): a logged-out LLM no longer crashes startup; failover skips it by @nicoloboschi in #6108
chore(deps): bump setuptools in root uv.lock by @cdbartholomew in #6091
chore(deps): bump fastmcp to 3.4.11 in root uv.lock by @cdbartholomew in #6113
fix(reranker): bench a failed failover-chain member so an outage does not slow every recall by @nicoloboschi in #6112
feat(decision-model): count the request ceiling and recall's caps in the model's own vocabulary by @nicoloboschi in #6115
feat(decision-model): one standalone file for recall's decision-rerank questions by @nicoloboschi in #6128
chore(reranker): deprecate the typesafe reranker provider in favour of the shared decision model by @nicoloboschi in #6130
system-evals: one file per operation, categorised cases, retain and recall capability suites by @nicoloboschi in #6118
fix(retain): store the 'where' dimension and fail instead of dropping an unsplittable chunk by @nicoloboschi in #6133
feat(recall): prune flag for the decision-model reranker by @nicoloboschi in #6131
fix(audit): an MCP tool call writes one audit row, not two (#6071) by @nicoloboschi in #6142
fix(recall): break BM25 score ties on id so repeated recalls return the same list by @nicoloboschi in #6143
blog: what's new in Hindsight Cloud, October 2026 by @benfrank241 in #6141
fix(refresh): bound the delta gate per refresh; a late failure keeps its memory by @nicoloboschi in #6153
fix(embed): keep the user's HINDSIGHT_API_DATABASE_URL instead of replacing it with pg0 by @nicoloboschi in #6158
fix(worker): shutdown no longer waits out a long maintenance job by @cdbartholomew in #6092
docs(dev): light start-docs that compiles pages on first visit (~1 GB instead of ~3.5 GB) by @nicoloboschi in #6159
fix(docker,helm): give the API 60s to hand back in-flight tasks on shutdown by @nicoloboschi in #6162
feat(recall): name the serving reranker in the trace (opt-in) and tag recall log lines with one id by @nicoloboschi in #6154
docs(coding-agents): explain how OpenCode 2.2 finds the v2 entry, pin exports['.'] to v1 by @nicoloboschi in #6170
docs(changelog): drop a non-change from coding-agents 0.8.3 by @nicoloboschi in #6173
feat(recall): live dedup flag for the decision-model reranker, prune and dedup per request by @nicoloboschi in #6139
fix(refresh): delta edits keep table rows, drop departed members, and refuse spilled operations by @nicoloboschi in #6171
fix(recall): the decision model sees each memory's date; recall eval replays frozen banks at scale by @nicoloboschi in #6168
fix(db): require asyncpg>=0.32 so a dropped connection can't wedge the worker by @nicoloboschi in #6107
chore(deps): bump authlib in google-adk uv.lock by @cdbartholomew in #6180
chore(deps): bump langgraph and langgraph-sdk in langgraph uv.lock by @cdbartholomew in #6182
chore(deps): bump qs and @hono/node-server in coding-agents by @cdbartholomew in #6183
chore(deps): bump rustls in Rust client Cargo.lock by @cdbartholomew in #6185
feat(decision-model): add a hosted provider option and a sign-in command by @nicoloboschi in #6169
docs: add OpenBot integration page by @benfrank241 in #6160
docs: add Hindsight Memory skill for OpenBot by @benfrank241 in #6161
feat(consolidation): one answer entry per fact, built for small models by @nicoloboschi in #6164
feat(llm): every tool call through a response schema; drop native tool calling by @nicoloboschi in #6176
hindsight provider: final model names (hs-decide-1-flash, hs-rerank-1, hs-embed-1) by @nicoloboschi in #6192
feat(mcp): recall takes rerank prune/dedup, like the HTTP API by @nicoloboschi in #6195
blog: OpenBot integration, shared memory over MCP by @benfrank241 in #6188
fix(embed): control center Unicode profiles, unhealthy-daemon stop, safe delete by @nicoloboschi in #6193
fix(coding-agents): pi custom-message runs no longer throw; auto-update reads npm 12 by @nicoloboschi in #6214
perf(lint): lint.sh 5-10x faster with the same checks; fix reflect done() crash (#6199) by @nicoloboschi in #6194
feat(llm-trace): opt-in flag to delete a document's request history with the document by @nicoloboschi in #6218
fix(reflect): an alias inside an identifier is not a citation (#6150) by @nicoloboschi in #6215
fix(anthropic): cache the prompt-injected schema in its own system block by @nicoloboschi in #6216
feat(retain): separate reasoning effort and call settings for the vision slot by @nicoloboschi in #6210
fix(metrics): count client-disconnect cancellations as success="cancelled" by @nicoloboschi in #6212
fix(llm): send images and files to Responses-API providers in Responses shape by @nicoloboschi in #6213
fix(startup): bound the startup LLM probe so a busy backend can't stall or fail boot by @nicoloboschi in #6219
chore(deps): bump qs in root package-lock.json by @cdbartholomew in #6184
chore(deps): bump litellm in litellm integration uv.lock by @cdbartholomew in #6181
fix(llm-trace): write out buffered trace rows before a document delete removes them by @nicoloboschi in #6221
chore(deps): bump vite in embed control center UI package-lock.json by @cdbartholomew in #6223
feat(llm-requests): show how many days are kept, hourly 24h chart, split by model or operation by @nicoloboschi in #6222
chore(deps): bump authlib and dnspython in root uv.lock by @cdbartholomew in #6179
fix: flaky tests — lock-order deadlocks, EvalPlanQual error, background-write races by @nicoloboschi in #6225
feat(embeddings): configurable extra body for OpenAI-compatible embedding requests by @nicoloboschi in #6227
fix(embeddings): run local embeddings on a fixed pool, not the default executor (#6110) by @nicoloboschi in #6226
fix(retain): read times with no UTC offset in the content's zone, not UTC (#6208) by @nicoloboschi in #6224
feat(knowledge): show the images behind a knowledge page's evidence by @nicoloboschi in #6217
fix(entities): keep mention_count exact under concurrent deletes and edits of one memory by @nicoloboschi in #6233
fix(consolidation): the decision model never writes observation text by @nicoloboschi in #6211
fix(reflect): every piece of evidence carries its dates, raw chunks included by @nicoloboschi in #6236
fix(reflect): the decision model's cut and verdict are retried on a rate limit by @nicoloboschi in #6237
perf(mental-models): let the memories store answer a model's staleness in one call by @nicoloboschi in #6261
fix(knowledge): a page's evidence shows its images on a store-owned bank by @nicoloboschi in #6262
fix(embeddings): fail loudly on a NaN/Inf embedding instead of at the pgvector bind by @nicoloboschi in #6268
fix(codex): send session-id so one operation's calls share the prompt cache by @nicoloboschi in #6267
fix(tracing): stamp hindsight.bank_id on every span in a bank operation by @nicoloboschi in #6269
fix(embed): find hindsight-api in the venv holding the package when another Python loads it (#6252) by @nicoloboschi in #6272
fix(consolidation): a batch whose LLM call fails writes no triage verdict (#6239) by @nicoloboschi in #6273
fix(transfer): export and import mental-model history a page at a time (#6259) by @nicoloboschi in #6274
fix(api): answer 4xx, not 500, for malformed ids, priorities and unstorable text by @nicoloboschi in #6276
fix(retain): a bank's store_document_text=false holds on delta retain and import by @nicoloboschi in #6270
feat(reranker): default to hindsight-ai/hs-rerank-1 by @nicoloboschi in #6263
fix(metrics): count a validator rejection as rejected, not as a failed operation by @cdbartholomew in #6256

Full Changelog: v1.1.0...v1.2.0