v0.14.2 - Service Automation Hardening

Agent Memory Bridge 0.14.2 is a conservative patch over the governed learning-candidate release. It keeps the 10-tool MCP surface stable while making the always-on service safer for multi-runtime use.

Thesis

0.14.2 = governance and embedding maintenance can run in the background without silently turning raw session/process chatter into durable memory.

What Changed

• Added explicit watcher.enabled and reflex.enabled service gates.
• Kept watcher/reflex disabled by default for normal always-on service use.
• Kept strong consolidation disabled by default and documented it as a deliberate replay/review workflow.
• Updated the service output so disabled watcher/reflex lanes report enabled = false, processed_count = 0, and reason = disabled.
• Clarified that AMB-native learning candidates, governance triggers, signals, and embedding maintenance do not require Codex rollout-log watching.
• Documented that multi-runtime installs should prefer AMB-native candidate/signal writes over client-specific log scraping.
• Added regression coverage for the conservative service defaults and disabled-lane behavior.
• Preserved the public MCP tool surface at 10 tools.

Evidence

Current release snapshot:

• pytest: 289 passed
• focused service/config subset: 45 targeted checks completed locally
• service --once: watcher, reflex, and consolidation report disabled when configured off; governance is idle without pending candidates; embedding sidecar is enabled but interval-gated
• index-health: FTS and embedding sidecars synchronized with zero missing, stale, or orphan rows in the local validation snapshot
• canonical lexical benchmark: question_count = 11, memory_expected_top1_accuracy = 1.0, memory_mrr = 1.0, file_scan_expected_top1_accuracy = 0.636
• procedure governance: governed_case_pass_rate = 1.0, governed_blocked_procedure_leak_rate = 0.0
• signal contention: signal_contention_case_pass_rate = 1.0, duplicate_active_claim_count = 0
• adversarial memory governance: 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 unchanged at 10 tools:

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

This release does not remove watcher/reflex/consolidation. It makes their always-on service use opt-in. Direct engine tests and deliberate replay/review workflows can still exercise those layers.

This release does not turn AMB into a scheduler, runtime agent, hosted memory platform, distributed lock, UI, or automatic durable writeback path from raw transcripts. FTS and embedding rows remain derived indexes, not durable authority.