vectorize-io/hindsight v0.8.0

on GitHub

What's Changed

• docs(guides): add Hermes memory guide batch by @benfrank241 in #1932
• docs: changelog and blog post for v0.7.2 by @nicoloboschi in #1933
• feat: upgrade MiniMax default model to M3 by @octo-patch in #1914
• chore(control-plane): bump next back to ^16.2.6 (undo 16.2.5 pin) by @nicoloboschi in #1934
• fix(retain): pre-extraction freshness recheck + serialize concurrent same-doc writers by @nicoloboschi in #1930
• blog: Using Entity Labels to Automatically Tag Memories in Hindsight by @benfrank241 in #1935
• feat(api): per-bank LLM request tracing via OTel GenAI recorder by @nicoloboschi in #1922
• fix(recall): gate VectorChord BM25 + add per-source candidate cap (#1707) by @nicoloboschi in #1947
• fix(config): persist bank config PATCH for never-retained banks (#1940) by @nicoloboschi in #1945
• fix(control-plane): localize operations and graph legends by @MapleEve in #1946
• fix(retain): stop bank_id routing key polluting fact attribution (#1680) by @nicoloboschi in #1948
• feat(transfer): export/import documents between banks without re-running the LLM by @nicoloboschi in #1909
• blog: Voice Agents That Remember — Adding Memory to Vapi with Hindsight by @benfrank241 in #1949
• feat(control-plane): "not enabled" splash for disabled audit logs & LLM requests (+ bank name fix) by @nicoloboschi in #1950
• fix(consolidation): eliminate duplicate observations (interleave dedup recall + tool & benchmark) by @nicoloboschi in #1907
• fix(retain): make document lock/upsert dialect-aware for Oracle (#1944) by @nicoloboschi in #1952
• fix(test): use current default model in Vertex AI integration test by @nicoloboschi in #1972
• feat(transfer): whole-bank export/import for cross-instance migration (#1884) by @nicoloboschi in #1953
• fix(search): escape pgroonga BM25 query text by @xmh1011 in #1966
• feat(recall): per-strategy retrieval boost via env config by @nicoloboschi in #1974
• feat(llm): provider prompt-prefix caching — retain + consolidation + reflect (bank-agnostic, default-on) by @cdbartholomew in #1936
• fix(autogen): default to Cloud + gated E2E + bucketing + ADD missing CI job by @DK09876 in #1868
• refactor(engine): cheaper bank stats — drop unused join, add freshness helper, result cache by @cdbartholomew in #1859
• blog: Long-Term Memory for Google ADK Agents with Hindsight by @benfrank241 in #1979
• blog(google-adk): update cover image by @benfrank241 in #1985
• fix(llamaindex): default to Cloud + replace dead manual test with gated E2E + requires_real_llm bucketing by @DK09876 in #1867
• feat(consolidation): semantic dedup of near-duplicate observations (create + update) by @nicoloboschi in #1977
• fix(clients): expose reflect tool_calls/llm_calls trace in python + typescript wrappers by @nicoloboschi in #1997
• fix(claude-code): default enableKnowledgeTools to true; keep MCP server alive when disabled by @nicoloboschi in #1999
• feat(api): enable LLM request tracing by default with 1-day retention by @nicoloboschi in #1996
• docs(configuration): document HINDSIGHT_API_BANK_STATS_CACHE_TTL_SECONDS + _MAX_ENTRIES from #1859 by @r266-tech in #1993
• fix: raise _check_health default timeout 2s→10s to stop busy-daemon kill loop by @21Felix04 in #1992
• docs(retrieval): correct "no per-strategy weight multipliers" after RECALL_STRATEGY_BOOSTS (#1974) by @r266-tech in #1991
• docs(configuration): document HINDSIGHT_API_LLM_PROMPT_CACHE_ENABLED in the LLM Provider table by @r266-tech in #1990
• docs(models): vertexai default model gemini-2.0-flash-001 is retired — sync to gemini-2.5-flash-lite by @r266-tech in #2001
• docs(api/bank-templates): fix entity_labels manifest example — label-group objects, not string[] by @r266-tech in #1984
• chore(deps): bump the uv group across 18 directories with 2 updates by @dependabot[bot] in #1982
• fix(mental-models): create bank before insert by @oldschoola in #1994
• docs(configuration): document HINDSIGHT_API_WORKER_IMPORT_DOCUMENTS_MAX_SLOTS worker slot reservation by @r266-tech in #1978
• docs(admin-cli): document decommission-worker --yes/-y confirmation-skip flag by @r266-tech in #1957
• fix(recall): bound temporal entry-point scan to top-50-per-fact_type (alternative to #1958) by @nicoloboschi in #1983
• feat(consolidation): enable observation dedup by default (0.97), skip on Oracle by @nicoloboschi in #2000
• fix: validate embedding dimensions before pgvector writes by @ai-ag2026 in #1670
• chore: update gemini embedding model from preview to GA by @dcbouius in #1780
• chore: update Gemini model recommendations to 3.x series by @dcbouius in #1787
• hindsight-docs: Added OutSystems community integration by @stefan-d-p in #1873
• docs(models): register fireworks so the Models grid + default-models table list it (#1860) by @r266-tech in #1911
• docs(cli): replace removed opinion fact-type with observation in recall example by @r266-tech in #1917
• docs(admin-cli): document full backup table coverage from #1903 by @r266-tech in #1929
• chore(deps): bump the npm_and_yarn group across 1 directory with 2 updates by @dependabot[bot] in #1938
• feat(embeddings): add ONNX local provider by @shoveller in #1970
• fix(oracle): make recall and mental-model history work on the Oracle backend by @DK09876 in #1980
• docs(performance): expand local-LLM concurrency guidance into a Local & Small Environments tuning section by @nicoloboschi in #2002
• feat(llm): apply HINDSIGHT_API_LLM_EXTRA_BODY across all API providers by @nicoloboschi in #2006
• feat(llm): opt-in strict json_schema across json_schema-capable providers (not just OpenAI) by @nicoloboschi in #2003
• fix(api): add vchord catalogs to search_path for external Postgres (#1351) by @nicoloboschi in #2008
• refactor(api): unify lazy bank-create into _ensure_bank_exists, couple to caller txn by @nicoloboschi in #2004
• fix(ci): repair zeroentropy embedding tests and regenerate drifted clients by @nicoloboschi in #2009
• fix(docker): clear diagnostic for pg0 bind-mount permission failure (#1483) by @nicoloboschi in #2010
• feat(history): move mental-model & observation history into dedicated tables by @nicoloboschi in #2007
• fix(reflect): let a fresh mental model short-circuit forced retrieval (no extra LLM call) by @nicoloboschi in #2011
• docs(models): note Groq free tier (8k TPM) is unsuitable for Hindsight by @nicoloboschi in #2015
• fix(init): fail fast when model init blocks instead of hanging forever by @nicoloboschi in #2014
• fix(llm): downgrade tool_choice="required" for servers that silently drop it (vLLM/LM Studio/Ollama) by @nicoloboschi in #2016
• feat(operations): durable progress snapshot for consolidation and batch retain by @nicoloboschi in #2013
• blog: How Hindsight Learns — A Deep Dive Into Mental Models by @benfrank241 in #2021
• LangGraph: add memory_instructions, fix nodes, remove BaseStore by @DK09876 in #1673
• feat(claude-agent-sdk): add Claude Agent SDK integration by @DK09876 in #1582
• fix(litellm): injection_mode, context manager restore, validation, error consistency by @DK09876 in #1711
• fix(openai-agents): default to Cloud + gated E2E + requires_real_llm bucketing by @DK09876 in #1866
• fix(opencode): default to Hindsight Cloud + gated live E2E by @DK09876 in #1915
• feat(integrations): add Superagent safety middleware by @DK09876 in #1128
• docs(api): document the operation progress snapshot + include_payload on the status endpoint by @r266-tech in #2037
• docs(claude-code): correct enableKnowledgeTools default (false→true) and disabled-behavior after #1999 by @r266-tech in #2044
• fix(reranker): make litellm-sdk reranker api_key optional for Bedrock IAM auth by @r266-tech in #2043
• fix(opencode): drop non-function export from plugin entry so it loads under OpenCode 1.16 (#2028) by @r266-tech in #2038
• fix(trace): preserve RRF source ranks by @zwcf5200 in #2040
• fix(transfer): include mental_model_history count in import-bank CLI summary by @r266-tech in #2032
• docs(models): flag Fireworks AI Batch API support in the provider capabilities table by @r266-tech in #2036
• docs(configuration): document shared cohere/litellm fallback API-key aliases by @r266-tech in #2031
• docs(models): sync gemini + vertexai default models to 3.x matching config.py by @r266-tech in #2030
• fix: remove Markdown bold formatting from fact extraction prompt by @Oxygen56 in #2029
• chore(deps): bump the uv group across 4 directories with 6 updates by @dependabot[bot] in #2027
• docs(models): document the onnx embeddings provider by @r266-tech in #2020
• fix(retain): expose retain outcome metadata by @xmh1011 in #2041
• docs(configuration): document HINDSIGHT_API_DATABASE_BACKEND (postgresql|oracle) by @r266-tech in #2024
• feat(recall): add configurable HINDSIGHT_API_SEMANTIC_MIN_SIMILARITY by @zwcf5200 in #2039
• feat(consolidation): periodic reconcile + cross-tenant retention via maintenance loop (#1969) by @nicoloboschi in #2019
• fix(opencode): observable logging — config-only debug, resolved-endpoint log, surfaced errors by @nicoloboschi in #2047
• perf(recall): add recall-temporal suite that forces the temporal arm by @nicoloboschi in #2046
• test(ci): de-flake TEI parallelism timing + disposition judge reruns by @nicoloboschi in #2045
• fix(opencode): call OpenCode app.log as a method so logging actually works by @nicoloboschi in #2049
• test(ci): harden LLM-as-judge against single-call verdict flips by @nicoloboschi in #2050
• docs(integrations): single source of truth (integrations.json) for gallery + sidebars by @nicoloboschi in #2048
• test(entity-labels): reproduce paired id/name extraction from [[...]] tags by @nicoloboschi in #2051
• fix(consolidation): set output token budget by @xmh1011 in #1967

New Contributors

• @21Felix04 made their first contribution in #1992
• @oldschoola made their first contribution in #1994
• @stefan-d-p made their first contribution in #1873
• @shoveller made their first contribution in #1970
• @Oxygen56 made their first contribution in #2029

Full Changelog: v0.7.2...v0.8.0