v0.2.9

Changed

• Calmer startup when on-device models aren't present. A semantic-capable build (the default release binary) running before pull-models now logs a calm NOTE and serves lexical BM25/FTS, instead of an ERROR plus a desktop alarm. Lexical-only is a supported default; the loud ERROR + alarm + refuse-to-serve is now reserved for MEMKEEPER_REQUIRE_SEMANTIC=1 (fail-closed).

Full changelog: https://github.com/teflon07/memkeeper/blob/main/CHANGELOG.md

v0.2.8

Changed

• Richer MCP tool definitions. All 16 tools exposed over memkeeper mcp now carry full, agent-facing descriptions: each states its purpose, when to use it versus similar tools, and whether it is read-only or mutating, and every parameter documents its type, default, and constraints. No change to the tool surface (names/arguments) — better guidance for agents and higher tool-definition-quality scores.

Full changelog: https://github.com/teflon07/memkeeper/blob/main/CHANGELOG.md

v0.2.7

Added

• doctor reports semantic readiness. memkeeper doctor now includes a semantic.models check: ok when the local embed model is present, a warning pointing at pull-models (and naming the resolved model dir) when a semantic-capable build has no models yet, or a note for lexical-only builds. Overall doctor status is unchanged — lexical still works — so it's guidance, not a failure.

Full changelog: https://github.com/teflon07/memkeeper/blob/main/CHANGELOG.md

v0.2.6

Release-first onboarding: the downloaded binary now delivers the on-device semantic experience with no rebuild.

Changed

• Release binaries are now self-contained and semantic-capable. The published macOS/Linux binaries bundle the ONNX runtime, so on-device semantic search works with no rebuild — run memkeeper pull-models once to fetch the models (lexical BM25/FTS until then). Previously the release binary was lexical-only and local semantic required building from source.
• Zero-config local models. memkeeper looks for the embed/rerank models in the pull-models default location ($MEMKEEPER_MODELS_DIR, else ~/.memkeeper/models) when the *_MODEL_DIR vars are unset — so pull-models then search turns on semantics with no environment variables. A missing model dir prints an actionable pull-models hint instead of silently degrading.

Added

• install.sh — one-line installer: downloads the release binary for your platform, verifies its SHA-256 (fail closed), installs to ~/.local/bin.

Full changelog: https://github.com/teflon07/memkeeper/blob/main/CHANGELOG.md

v0.2.5

Added

• Alias-exact-match retrieval boost. A query token that exactly matches a memory's reserved alias::<normalized> tag now adds a fixed boost to that memory's retrieval score, lifting exact alias hits (e.g. "k8s" → the Kubernetes memory) above semantically-similar neighbors near the abstention floor. Matching uses normalized 1–3-word query shingles, so multi-word aliases ("dead letter queue") resolve too. Reuses tag storage; no schema or wire-protocol change.

Full changelog: https://github.com/teflon07/memkeeper/blob/main/CHANGELOG.md

v0.2.4

memkeeper v0.2.4

v0.2.3

memkeeper v0.2.3: --json @file/stdin, serve --store, shell-aware pull…

v0.2.2

memkeeper v0.2.2: experimental Windows docs+CI, onboarding docs, dash…

v0.2.1

memkeeper v0.2.1: macOS + Linux binaries; Windows in progress

v0.2.0

memkeeper v0.2.0: off-device semantic, three retrieval modes, Windows…