Skip to content

1.1.0 - 2026-06-15

Choose a tag to compare

View all tags
@github-actions github-actions released this 15 Jun 23:59
v1.1.0
3db1b10
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Release Notes

Makes the warm worker's index repair non-blocking, cuts HNSW cold-start cost,
and adds embedding-device pinning and a prebuilt-first installer.

Added

  • MEMD_EMBED_DEVICE: pin the embedding device (cpu, cuda, cuda:N)
    instead of always taking cuda:0, so memd can stay off a contended GPU on
    shared machines.
  • Prebuilt-first make install: installs the released binary when it runs
    on the host and compiles from source only as a fallback, so a first install
    no longer requires a Rust toolchain.
  • warm status repair signal: the ryw_probe payload now reports
    repair_in_progress, so an in-flight background index repair is observable.

Changed

  • Non-blocking external-mutation repair: when the warm worker detects an
    external metadata.db mutation it now schedules a single-flight background
    HNSW repair and serves the request after a short bounded wait, instead of
    running the full backfill synchronously in the request path. Previously a
    detected mutation could block a warm request for minutes and trip the client
    timeout (notably on shared/NFS data directories). The worker still indexes
    its own warm-routed writes synchronously, so same-worker read-your-writes is
    unchanged; dense/hybrid coverage of externally-added chunks catches up once
    the background repair lands (SQLite and sparse reads are immediate).
  • Honest warm client-timeout message: the timeout no longer asserts a
    successful cold-path fallback that may not have happened; it points at the
    worker log and memd warm status.

Fixed

  • add_batch usage events: batched adds now record per-chunk usage events
    (mirroring single add), so the usefulness report and the central per-chunk
    hit log no longer undercount batched writes.

Performance

  • Cheap HNSW cold-start: the cold-start backfill reuses persisted
    embeddings (embeddings.bin) and re-embeds only the genuine delta via a
    cache-aware membership check, instead of re-embedding every chunk from text.
    A clean restart on a 140-chunk tenant drops from ~50 s to no re-embeds.

Install memd 1.1.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/fmschulz/memd/releases/download/v1.1.0/memd-installer.sh | sh

Download memd 1.1.0

File Platform Checksum
memd-aarch64-apple-darwin.tar.xz Apple Silicon macOS checksum
memd-x86_64-apple-darwin.tar.xz Intel macOS checksum
memd-aarch64-unknown-linux-musl.tar.xz ARM64 MUSL Linux checksum
memd-x86_64-unknown-linux-musl.tar.xz x64 MUSL Linux checksum
Assets 15
Loading