Memory layers and multimodal inputs are finally moving from novelty to real-world apps. The hottest chatter centers on an AI App Store where apps can share a user-controlled memory layer, letting Travel Planner, Packing Assistant, and Budget Planner pre‑remember your preferences—only with your OK to share. There’s also a built-in My Context dashboard you own. [1]
That setup smooths onboarding and keeps you in control. You can edit or delete memories, and apps request access rather than hoard data. It’s early, but the pattern points to memory as a first-class feature in usable AI. [1]
Modalities and performance — On the multimodal front, a thread compares Qwen3VL against Qwen3 and Qwen3-VL-30B-A3B-Instruct. People report that adding a visual module can muddy pure text performance, with some noting text tasks degrade while multimodal tasks gain from visual-spatial cues. [2]
Hellocafe.ai — In voice demos, Hellocafe.ai shows a cafe ordering flow with 100% local models, using Llama 8B and Whisper, deployed on Kubernetes. [3]
For voice agents, evaluators lean on dedicated tools: - Deepgram Eval – transcription accuracy testing [4] - Speechmatics – multilingual evaluation [4] - Voiceflow Testing – end-to-end dialogue flows [4] - Play.h.t Voice QA – TTS voice quality [4] - Maxim AI – structured end-to-end voice eval (latency, persona tests) [4] - Eleven Labs – strong contender for voice evaluation [4]
On the knowledge base side, the talk is about KB quality in RAG: contradictions, duplicates, freshness, coverage, readability, safety and provenance. Some propose a knowledge graph to surface issues automatically. [5]
Memory, modality, and KB health are shaping practical LLM apps you can actually use—watch how memory ownership, cross-modal tradeoffs, and evaluation tooling evolve next. [1][2][5]
References
Show HN: AI App Store with a User-Controlled Shared Memory Layer
Shows AI App Store enabling user-controlled shared memory between apps; seeks feedback on memory model, dashboard, and features so far.
View sourceFrom your experience for text only, how is Qwen3VL compared to Qwen3, does having a Visual module penalize the text-only capacities ?
Compares Qwen3-VL versus Qwen3 text-only; notes text degradation with vision, mixed math performance, and mmproj implementation considerations in practice.
View sourceShow HN: Hellocafe.ai – open-source voice and chat enabled AI ordering system
Open-source demo shows local LLM (Llama 8B) and Whisper for chat/voice ordering, Kubernetes deployment, base for AI apps.
View sourceThe best tools I’ve found for evaluating AI voice agents
Discusses tools and metrics for evaluating AI voice agents integrating STT, LLMs, and TTS across accents and multi-turns
View sourceHow do you evaluate the quality of your knowledge base?
Proposes metrics to assess knowledge base quality in RAG for LLMs; suggests knowledge graph to reveal duplicates, contradictions, and gaps.
View source