1.6.0 - 2026-07-25

Release Notes

Changed

• OutcomeKind is now #[non_exhaustive]. Adding verifier_error below breaks
  any downstream exhaustive match on it, so the marker lands in the same
  release: consumers absorb one break now instead of another every time the
  eligibility model gains a state. Match arms on the enum need a _ arm.

Added

• verifier_error outcome state, for a verifier that produced no verdict
  because it crashed, timed out, or returned something unparseable. Until now
  the closest available state was failed, which asserts the task was verified
  and did not succeed and therefore penalises the chunks that were retrieved.
  Reporting a broken verifier that way turns endpoint flakiness into permanent
  negative priors on memories that may be perfectly good. Observed at benchmark
  scale while judging a memory benchmark: 41 empty judge responses and 19
  client timeouts were scored as wrong answers for the system under test.

  A verifier_error credits nothing in either direction, even when used and
  harmful chunk IDs are supplied, so it can be reported without knowing what
  the run would have concluded. The event is still persisted and queryable, so
  verifier failures stay visible instead of silently moving ranking. The
  evidence requirement is unchanged: verifier types that require an evidence
  reference still require one, since a broken run is exactly when the log or
  job reference matters.

  No migration is needed, but the state is not readable by older binaries.
  Downgrading after recording one makes the outcome list for that episode fail
  to parse; other episodes and all ranking reads are unaffected.

Fixed

• Background consolidation had no machine-wide ceiling. The 1.5.1 lock is keyed
  on tenant and project, which stops one project stacking spawns but leaves the
  machine unbounded: every project whose dirty region crosses MIN_REGION adds
  another detached child, and they all queue on the one store writer. A machine
  with several active projects could therefore still reproduce the starvation
  1.5.1 set out to fix. Spawning now also claims one of four interchangeable
  slot locks under the data directory, so at most four background
  consolidations run at once regardless of how many scopes are active. A spawn
  past the ceiling returns {"skipped": "consolidation_slots_busy"} and is
  reported by session-start as a skip reason rather than a start. The scope
  lock is still claimed first, so a contended scope reports itself as such
  instead of consuming a slot it would not use.

• session-start dropped background-spawn errors. Only a skipped reason was
  surfaced, so a spawn that failed outright reported
  consolidation_spawned: false with no reason, which reads as "nothing needed
  doing". 1.5.1 moved scope resolution and lock acquisition into the parent,
  making that path reachable, and the error is now reported.

• cargo test was intermittently red, roughly three runs in twenty. The 1.5.1
  lock test holds a child process while the rest of the suite runs, and fork
  copies the whole descriptor table, so an unrelated test asserting on a freshly
  released file lock could observe a copy still held by a child that had not yet
  reached its exec. Both affected assertions now retry briefly, matching the
  allowance the 1.5.1 test already made for its own release check. Twenty
  consecutive runs pass.

Install memd 1.6.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/fmschulz/memd/releases/download/v1.6.0/memd-installer.sh | sh

Download memd 1.6.0

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