2026.4.13 — April 2026 — Loquent Changelog

This release makes the AI assistant dramatically more reliable for bulk operations, syncs large call histories in the background without blocking the UI, and restructures AI prompts for ~80% cache hit rates — cutting latency and cost per conversation turn. 6 PRs merged.

Operation Scratchpad

The AI assistant now tracks progress on bulk operations with persistent memory that survives across conversation turns.

Persistent bulk tracking — when you ask the assistant to work through a large set of contacts (e.g., “message 50 contacts tagged Lead”), it writes progress to a scratchpad that persists beyond the conversation window
Resume where you left off — if the assistant’s context resets mid-operation, it picks up exactly where it stopped — no duplicate messages, no missed contacts
Bulk contact tools — new tools like get_bulk_contact_memory and get_bulk_contact_messages let the assistant fetch enriched contact data in batches, with memory snippets and message previews included
Hardened reliability — org-scoped filtering, deduplication, status tracking per contact, and optimized batch queries ensure operations complete cleanly at scale

Background Call History Sync

Syncing phone numbers with thousands of calls is now fast and non-blocking.

Instant response — triggering a call history sync returns immediately instead of waiting for the full download
Page-at-a-time processing — calls are fetched and stored incrementally with real-time progress updates, keeping memory bounded regardless of volume
Stale recording recovery — a lease-based system prevents duplicate processing of call recordings and safely reclaims stalled claims

Prompt Cache Optimization

AI conversations are now faster and cheaper thanks to restructured system prompts.

Static/dynamic split — system prompts are reorganized with static content first and dynamic context last, maximizing cache reuse across turns
~80% cache hit rate — most of the prompt is served from cache on subsequent turns, reducing latency and token costs
~400 fewer tokens — redundant context blocks removed without losing any capability

Infrastructure & Reliability

Migration timestamp fix — corrected a migration filename to match the name already applied to staging, preventing server errors on deploy
Notification title truncation — notification titles are now safely truncated to fit database constraints, preventing write failures on long titles