1.5.0 - 2026-07-22

Release Notes

Fixed

• memd call and memd batch now inherit missing tenant_id and
  project_id values from .memd/project_scope.json, including requests
  routed through a warm worker. An explicit JSON tenant_id remains
  tenant-wide unless the same request also supplies project_id. Malformed
  or unreadable scope files fail closed, and batch scope failures preserve
  per-line receipts under --continue-on-error.

• bm25-only no longer loads the dense index or applies recency, project,
  type, or query-text feature bonuses. It preserves BM25 order before any
  stored-feedback adjustment.

• Sparse segment compaction now waits for an explicit Tantivy merge instead of
  issuing a no-op commit. The compaction runner and
  memd maintenance --aggressive now perform the merge; the maintenance
  report includes sparse segment counts before and after it.

• memory.add and memd add return every physical chunk ID created by
  long-document splitting in stored_chunk_ids, with the backward-compatible
  chunk_id first. memory.supersede returns the replacement set in
  new_stored_chunk_ids. Initial lifecycle metadata is applied to every split
  child, and failed supersession links compensate all children instead of
  deleting only the primary chunk.

• Fixed-clock searches ignore future-dated feedback and outcome events. Tier
  diagnostics now use the same metadata reranker as standard search, so debug
  mode no longer reports a different order from the served path.

• Tiered semantic-cache hits now run and fuse the sparse retrieval leg instead
  of returning a dense-only cached ranking. A regression test checks that cold
  and warm hybrid searches preserve the same fused result order.

• Exact lexical rescue now ranks the full bounded scan before retaining the
  best candidates, so earlier low-scoring matches cannot hide later correction
  records. Boundary punctuation is normalized without short-token substring
  matches, while dotted identifiers and natural inflections remain searchable.

• memory.md now deduplicates project and machine-wide candidates as one
  bounded union, including exact chunk IDs, consolidation lineage, and topic
  matches. Active-project results are assigned to the project section instead
  of appearing twice or leaking into the machine-wide section.

• Startup answerability no longer treats a missing or unreadable
  tasks/todo.md as proof that no tasks are open. The generated project state
  records the task source as missing, parsed-empty, parsed-open, or failed.

• CLI, structured-operation, batch, OMF-import, supersession, episode, and
  consolidation writes now share one preparation service for tag
  normalization, admission, retention, priority, and initial trust. Batch and
  single writes no longer diverge on inferred priority or retention defaults.

• Consolidation now admits chunks retrieved after the last-run watermark,
  even when those chunks were created earlier. Every candidate is checked
  against its current lifecycle and project scope before use.

• Tenant-wide consolidation no longer hides project-scoped source memories.
  It records derives_from:<csv> lineage and leaves those sources active;
  project-scoped consolidation retains its existing supersedes:<csv> and
  soft-tombstone behavior.

• WAL replay preserves newer SQLite lifecycle metadata when repairing an
  existing payload location, so restart cannot revert Final, Superseded,
  Expired, or Error rows to the status stored in their original Add record.

Added

• memory.search accepts ranking_time_ms to pin recency, feedback, and
  outcome decay over a frozen corpus. Fixed-clock searches are read-only with
  respect to the usage ledger and retrieval-attribution episodes. Lifecycle
  visibility remains live, so this option is not an as-of snapshot.

• Search responses always serialize retrieval_episode_id. Read-only
  fixed-clock searches return JSON null, allowing callers to reject binaries
  that silently ignore ranking_time_ms.

• Search and agent context now return privacy-safe retrieval episode IDs.
  memd outcome and memory.record_outcome accept explicit used/harmful
  attribution and verifier evidence. outcome-v1 computes bounded,
  project-scoped, time-decayed shadow ranks while production serving remains
  unchanged.

• memd eval-outcome-ranking writes JSON and Markdown reports comparing
  served and source-deduplicated shadow top-k lists against explicit relevant
  and harmful chunk judgments.

• Consolidation runs use durable run, entry, and normalized lineage tables.
  Outputs are written as hidden Candidate chunks. The default
  memd consolidate stops after validation; memd consolidate-review --list
  discovers staged runs, and --accept records durable promotion intent
  before atomically promoting candidates with their same-project sources.
  memd consolidate --promote explicitly requests that path in one command.
  Exact source-set retries reuse the existing run.

• LLM consolidation output must include a concrete agent action, exact source
  evidence, and bounded confidence. The journal records the consolidator
  command, model, and version and references a permission-restricted,
  size-capped raw-response audit artifact with integrity hashes.

• Session start performs bounded consolidation recovery before refreshing
  agent context. Recovery isolates per-run errors, protects fresh in-flight
  work with a 30-second grace period, promotes only runs with recorded intent,
  and refreshes promoted dense-index rows. Background consolidation stages
  proposals for later review instead of changing source visibility.

Changed

• Upgrading an existing store is conservative. Previously validated runs gain
  promotion_requested = 0 and appear in the review list, but cannot be
  promoted because they lack the new audit artifact; an acceptance attempt
  fails closed and rejects their candidates. Older planned or
  candidate-written runs without that artifact are rejected during recovery.
  Sources remain active in both cases. --legacy-immediate preserves the old
  one-command behavior for one release and prints a deprecation warning.

Install memd 1.5.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/fmschulz/memd/releases/download/v1.5.0/memd-installer.sh | sh

Download memd 1.5.0

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