v0.3.2 — notes are now hybrid-searchable

Fixed

• Notes added via memory.add_note (MCP tool or the gateway's /add-note) are now indexed into LanceDB + tantivy through the same capture → chunk → embed → index path the daemon uses, so /search hybrid retrieval finds them. Notes get an OLTP capture row, KG DERIVED_FROM provenance (so memory.forget --capture-id cascades), and an audit entry. If the embedder is unavailable, notes degrade to BM25-only indexing and the next secondbrain index bulk pass backfills vectors. (#7)
• The actants (Ollama) embedding backend no longer crashes when its sync surface is called from a running event loop (gateway handlers, daemon consume loop).

Install: pip install -U secondbrain-ai

Full changelog: https://github.com/openintelligence-labs/secondbrain/blob/main/CHANGELOG.md

SecondBrain v0.3.1

Now on PyPI: pip install secondbrain-ai (CLI and import name remain secondbrain). Also: local-model test defaults (no egress), configurable daemon LLM timeout, mcp<2 pin. See CHANGELOG.

v0.3.0 — redaction gate + audit-logged redactions

Local-first personal memory daemon — this release lands the sensitive-content redaction gate and makes the heavy ML stacks opt-in.

Added

• Redaction gate wired into the dedup cascade (heuristic baseline; the Florence-backed model lands behind a [redact] extra later).
• --redact / --redact-threshold flags on secondbrain run-synthetic for display-free smoke testing.
• Redaction events are recorded in the signed audit log, keeping the compliance trail complete when frames are dropped before persistence.
• [ml] and [visual] extras — sentence-transformers/torch and ColQwen2.5 are now opt-in instead of mandatory.
• mcp + macOS PyObjC moved into base dependencies so a clean install works on a fresh Mac.

Fixed

• Test-order pollution: a daemon started with enable_llm=True leaked its LLM-backed extractors into later tests; globals now reset via a shared conftest fixture.
• Two CI-runner-only test failures and all ruff lint/format errors.

Full test suite: 207 passed (incl. live-Ollama and macOS TCC tests). End-to-end verified against Ollama gemma4:latest: synthetic capture daemon with LLM importance + commitment extraction in the loop, HTTP gateway /health, /add-note, /search, /commitments, /audit-log.

As always: captures, embeddings, knowledge graph, and audit log stay on-device. Local-by-default, BYO-LLM.