v0.14.0 - Governed Learning Candidates

Agent Memory Bridge 0.14.0 adds a governed staging lane for runtime learning candidates while keeping the public MCP surface small and stable.

Thesis

0.14.0 = policy-gated learning candidates before durable memory writeback.

This release adds the missing boundary between a runtime noticing something useful and AMB treating it as ordinary durable memory. Runtime or harness layers may propose learning candidates, but AMB owns policy evaluation, staging, suppression, and later review/promotion boundaries.

What Changed

• Added a v0 learning-candidate policy evaluator with explicit writeback decisions:
  • allow
  • needs_review
  • deny
  • degraded_no_write
• Added internal learning-candidate storage with candidate statuses:
  • pending
  • needs_review
  • approved
  • rejected
  • expired
• Recomputed learning policy at the storage boundary so callers cannot forge an allow decision.
• Stored candidate records as review-stage memory records tagged with kind:learning-candidate, candidate_status:*, authority_class:*, source_runtime:*, and schema tags.
• Hid learning candidates from ordinary recall, browse, export, and stats unless explicitly requested through review tags such as kind:learning-candidate or candidate_status:*.
• Documented the learning-candidate review queue in the authority contract.
• Kept the public MCP tool surface unchanged at 10 tools. There is no new learning_candidate, startup_packet, or task_packet MCP tool.

Evidence

Current release snapshot:

• pytest: 228 passed
• public MCP tools: 10
• deterministic proof: 4/4 checks
• learning-candidate tests cover:
  • safe candidate allow decisions
  • review routing for high-authority classes
  • denial for malformed, sensitive, raw-transcript, and unsupported candidates
  • degraded no-write behavior when AMB is unavailable
  • storage rejection for forged policy decisions
  • suppression from normal recall, browse, export, and stats
• signal contention benchmark:
  • signal_contention_case_count = 5
  • signal_contention_case_pass_rate = 1.0
  • unique_active_claim_rate = 1.0
  • duplicate_active_claim_count = 0
  • active_reclaim_block_rate = 1.0
  • stale_ack_blocked_rate = 1.0
  • stale_reclaim_success_rate = 1.0
  • pending_under_pressure_claim_rate = 1.0
  • initial_hard_expiry_cap_rate = 1.0
• adversarial memory-governance benchmark:
  • adversarial_case_count = 6
  • adversarial_task_count = 7
  • adversarial_governed_task_pass_rate = 1.0
  • adversarial_governed_blocked_record_leak_rate = 0.0

Boundaries

The public MCP surface remains the same 10 tools:

• store, recall, browse, stats
• forget, promote, export
• claim_signal, extend_signal_lease, ack_signal

Learning candidates are staging records, not approved durable memory. A candidate can help a reviewer decide what to promote later, but it is not source-of-truth memory until a reviewed promotion or replacement path creates the final record.

This release does not add:

• a scheduler
• a watcher runtime
• a hosted backend
• an autonomous task runner
• a graph database
• a distributed queue
• a complete candidate review UI
• automatic durable writeback from raw transcripts

Next

Use 0.14.x for narrow release-polish work: docs alignment, small compatibility fixes, and review/promote ergonomics that do not expand the public MCP surface.

A future minor release should be reserved for a larger thesis such as pre-compaction capture, broader retrieval fixtures, or a complete reviewed candidate-promotion workflow.