QVerisBot 2026.3.8 Release Notes

QVerisBot 2026.3.8 release notes — memory upgrades, model switching, Qveris routing improvements, and a 60+ bugfix merge from OpenClaw upstream.

Previous version:2026.3.3 |Release date:March 8, 2026

This release merges the latest OpenClaw upstream (through v2026.3.7 stable) and introduces three major QVerisBot-specific enhancements.

New: Memory Enhancement

QVerisBot now ships with an advanced memory management system built on two bundled hooks:

• context-digest— Maintains a rolling cross-session digest that survives compaction and session resets, giving the agent persistent awareness of prior conversations without bloating the context window.
• session-importance — A two-stage importance classifier that evaluates conversation turns and prioritizes high-signal content for long-term retention.
• Shared infrastructure — New shared utilities (transcript-reader, llm-memory-helpers) provide transcript parsing, LLM orchestration with mutex/dedup, and a system prompt anchor (context-digest-anchor) that injects digest context at the right position in the prompt. Native memory flush is handled automatically at the end of each agent run.

New: Natural Language Model Switching

A new switch_model tool allows users to change the active LLM model through natural language — no config editing required.

• Fuzzy matching — Handles partial names, common aliases, and case-insensitive input (e.g., "use claude sonnet", "switch to gpt-4o").
• Ambiguity detection — When multiple models match a query, the agent presents candidates and asks the user to clarify instead of guessing.
• Silent reset notifications — When a model override is cleared (e.g., by session reset), a system event notifies the agent so it can inform the user transparently.

New: Qveris Integration Enhancement

The Qveris tool integration has been significantly upgraded with smarter routing and session awareness:

• Structured routing decision tree — buildQverisSection() now generates a 6-step decision tree in the system prompt with explicit anti-patterns (local filesystem ops, docs/tutorials, non-English queries), guiding the agent on when to use Qveris vs local tools vs web_search.
• **qveris_get_by_ids**tool — A new tool that verifies known tool IDs via POST /tools/get-by-ids without a full search, reducing unnecessary API calls when the agent already knows which tools to use.
• Session-scoped tool rolodex — Successful tool executions are recorded in a per-session rolodex. Search results are annotated with previously_used and session_uses metadata, and session_known_tools is exposed to the agent so it can reuse proven tools efficiently.
• Improved search boundaries — qveris_search descriptions now include negative boundaries and GOOD/BAD examples to prevent task-goal searches and improve search precision.

Merged from OpenClaw Upstream (v2026.3.7)

Key upstream changes included in this release:

Features

• Context Engine plugin interface with full lifecycle hooks for alternative context management strategies
• ACP persistent channel bindings for Discord and Telegram (survives restarts)
• Telegram per-topic agentId overrides for forum groups and DM topics
• Google Gemini 3.1 Flash-Lite first-class support
• Mattermost interactive model picker
• Gateway SecretRef support for gateway.auth.token
• Docker multi-stage build with slim variant support
• Compaction safeguard tuning and post-context configurability
• iOS App Store Connect release preparation
• Web search provider selection in onboarding wizard

Fixes

• 60+ bug fixes across gateway auth, Telegram routing, Discord session keys, Slack typing/dedup, Feishu streaming, cron delivery, sandbox hardening, voice-call config, TUI session isolation, and more
• Security hardening: cross-origin redirect header filtering, fs-bridge path safety, zip extraction writes, cron file permissions
• Dependency security patches for Hono and tar