Skip to content

ai-memory v0.3.1

Choose a tag to compare

@github-actions github-actions released this 24 Oct 14:43
a7947d2

Reviewed ai-memory release from a7947d2.

Changes since 0.2.8

The core npm package and paired GitHub plugin archive are both version 0.3.1. Upgrade directly from 0.2.8; 0.3.0 can be skipped. This release includes every change from 0.3.0 and the subsequent brief, security, tools and staleness work.

  • Explicit caller-session identity governs orientation recovery; a project fallback does not substitute another session's identity. Continuity checkpoints are excluded from durable knowledge retrieval, and indexed tag aggregates supply retrieval priors.
  • Secret-scan diagnostics expose value-free finding metadata while keeping secret values redacted (#67).
  • Curated knowledge retention: memory_flush saves continuity checkpoints without creating memories or embeddings. Backup-gated cleanup removes historical flush fragments, promoted session summaries and unused history while preserving curated knowledge, required lineage and contested pairs (#63).
  • Index-first search reads bounded candidates through text, key, tag and reference indexes before ranking their union (#66).
  • Scoped pinned project briefs appear at SessionStart, with role state first, explicit audience matching and a private cache per project/host/role. AI_MEMORY_ROLE selects the session role; ai-memory brief --out provides a context file when native hooks are unavailable (#77).
  • HTTP MCP rejects DNS-rebinding Host headers and creates a separate transport session for each client. Origin-bearing requests require an exact entry in AI_MEMORY_MCP_HTTP_ALLOWED_ORIGINS; the default allowlist is empty (#75).
  • Session-end hooks preserve the producing host and session identity, including Claude Stop payloads, rather than inventing a replacement (#76).
  • Memory tools support project-scoped keys and topic recall, with aligned clear/compact hook behavior (#80).
  • Staleness reports flag old verification and closed GitHub references; health reports show per-key injection/use evidence and prune candidates. Re-storing unchanged knowledge refreshes verification without creating a version (#79).
  • Disaster-recovery test fixtures recover owned descendant processes when PID readiness is incomplete; this changes test cleanup, not the backup format (#81).
  • Plugin installation registers cancellation handlers before launching the native host CLI, including while process-group ownership is recorded, so an early cancellation still cleans up the owned process group (#84).

Migrations since 0.2.8

0.2.8 already contains migrations 001–009, ending in 009_workflow_telemetry. The 0.3.1 initializer validates that existing prefix and applies these two pending migrations in order:

  1. 010_index_first_search builds the stored-vector GIN and corrected author/category indexes, merges and disables the tags pending list, then removes obsolete expression indexes. Index builds block writes; final index changes hold exclusive locks until commit. Stop older clients before administrator initialization.
  2. 011_pinned_brief adds audience (default all), pinned (default false) and last_verified_at (backfilled from each row's unchanged updated_at), plus the brief selection index and insert trigger. Existing memory IDs, content and other stored fields remain readable and unchanged.

No intermediate package installation is needed. Older initializers reject the longer ledger. Rollback requires a verified pre-upgrade backup restored into a new database with its matching old package and original migration files; changing registrations alone does not undo migrations.

Follow Adopting 0.3.1 for preserved registrations, paused writers, backups, migrations, role/brief/origin settings, cleanup and recovery. Publication does not run host adoption. Installed corpus, latency and seven-day acceptance remain separate operator work.