v3.0.0-beta.4

Completes Phase 2 of the ROADMAP. Adds domain partitioning, cold-tier
recovery, hard memory budgets, and adaptive scheduling hints — plus the
first release publishing via npm Trusted Publishing (OIDC) instead
of a long-lived token.

Added

• Memory namespaces (#16) — optional namespace argument on
  add / query / consolidate / timeline / stats / resume /
  export / import / pruneColdTier. Default namespace is
  'default', so existing callers are unaffected. Entities and
  relationships remain agent-scoped (knowledge graph shared across
  namespaces by design). Migration schema/migration-v3.1.sql
  adds a namespace TEXT NOT NULL DEFAULT 'default' column to
  eight memory tables with composite (agent_id, namespace) indexes.
  Public surface covers HTTP, MCP, TypeScript SDK, Python SDK, and
  OpenAPI. Sleep cycles are intentionally agent-wide (not
  namespace-scoped) — the 10-phase cycle doesn't yet filter by
  namespace, and SleepSchema explicitly does not accept a
  namespace argument so the API doesn't claim behavior it can't
  deliver.
• Cold tier search + restoration (#14) —
  MemoryManager.searchColdTier(agentId, opts) filters archived
  rows by content substring, namespace, archive-timestamp range,
  and source table with limit/offset paging.
  restoreColdTier(agentId, coldTierId, opts?) copies a cold row
  back into warm tier non-destructively (cold row preserved, warm
  row stamped with metadata._restored_from_cold_id for audit).
  New routes GET /memory/:id/cold and POST /memory/:id/restore,
  plus MCP tools and SDK methods. No schema migration needed.
• Memory budgeting — new opt-in WARM_TIER_MAX_PER_AGENT env
  var enforces a per-agent hard cap on warm_tier rows. Capacity
  eviction runs as Phase 2b of the sleep cycle (after threshold
  eviction) and archives lowest-importance rows — not oldest —
  until the cap is met. Uses the existing warm_tier_importance_idx
  for efficient scans. Graduated memories are NOT exempt: cap is a
  hard limit; graduation affects retrieval scoring only.
  SleepCycleResult.capacity_evicted is set when a cap is
  configured.
• Adaptive sleep advisory (Phase 2 adaptive scheduling) —
  MemoryManager.sleepAdvisory(agentId) returns a
  health-metric-based scheduling recommendation over five signals:
  hot backlog, contradiction rate, revision debt, time since last
  sleep, and an inverse stability signal that clamps urgency
  downward when the knowledge base is highly graduated.
  recommended flips to true at urgency 'medium' or 'high'.
  External orchestrators (cron, dashboards, control planes) call
  this instead of scheduling blindly — MemForge stays
  scheduler-free by design. New HTTP route
  GET /memory/:id/sleep/advisory, MCP tool memforge_sleep_advisory,
  and SDK methods on both TS and Python clients. Thresholds
  configurable via SLEEP_ADVISORY_* env vars.
• Optional HNSW index template (#95) — new
  schema/hnsw-indexes.example.sql documents how operators who use
  embeddings should pin the warm_tier.embedding and
  shared_memories.embedding halfvec columns to their provider's
  dimension and build HNSW indexes for fast cosine similarity.
  Common dimensions (384, 768, 1536, 3072) listed inline.

Changed

• npm publish switched to Trusted Publishing (OIDC) — release
  workflow no longer requires NODE_AUTH_TOKEN; the
  id-token: write permission is enough, and npm validates the
  publish came from this specific repo + workflow via OIDC. No
  long-lived token to rotate or leak. The NPM_TOKEN secret can
  now be deleted from the repo (kept as fallback for beta.4 out
  of caution).
• schema/schema.sql no longer tries to build HNSW indexes on
  bare-dimension halfvec columns (#95). The two failing
  CREATE INDEX ... USING hnsw statements have been replaced with
  comments pointing to hnsw-indexes.example.sql. Fresh installs
  without embeddings are now silent (no psql ERROR lines); installs
  with embeddings apply the example after choosing a dimension.
  Semantic search falls back to sequential scan on unindexed halfvec
  columns — correct but slower.
• Bumped package.json to 3.0.0-beta.4.
• Closed GitHub issues — #10 (npm publish), #14 (cold tier
  search + restoration), #16 (namespaces), #17 (per-agent weights
  — previously shipped in v2.4), #18 (CORS), #35 (audit
  immutability), #43 (global rate limiting), #46 (deployment
  security guide), #50 (knowledge graduation). All were already
  shipped but not yet closed on the tracker.

Known gaps documented

• #13 Multi-model revision — two-pass triage variant — still
  future work. The basic cheap/capable split via
  REVISION_LLM_PROVIDER already works today.
• #41 Redis cache poisoning — confirmed src/cache.ts does not
  Zod-validate deserialised cache entries. Future security sprint.
• #21 Integration guides — LangChain and CrewAI are covered;
  AutoGen remains gap in INTEGRATION.md.
• #22 Architecture Decision Records — still future docs work.