One-Line Summary

China’s Alibaba claims the viral top video model as new work on agent memory pushes reliability and observability to the forefront.

LLM & SOTA Models

Alibaba confirms HappyHorse-1.0 as the top anonymous video model

A fast-rising video generator called HappyHorse-1.0 shows up on a public leaderboard and quickly takes the top spot for both text-to-video and image-to-video; Alibaba now confirms it built it. The team reveals on X that HappyHorse comes from Alibaba’s ATH AI Innovation Unit, and Alibaba tells CNBC the post is genuine, ending days of speculation over whether a big tech firm or an indie created it. Alibaba’s Hong Kong shares close up 2.12% on the day of the reveal. ¹

The model first appears around April 7 on Artificial Analysis, a benchmark that ranks video models using head-to-head, human-preference-based comparisons. HappyHorse climbs to No. 1 across both categories, ahead of ByteDance’s Seedance 2.0 and Kuaishou’s Kling. SCMP reports the unit sits within the newly formed Alibaba Token Hub (ATH) and that more releases are in the pipeline. ²

The timing matters: OpenAI discontinues its Sora app and platform amid high compute costs and a shift toward coding tools and enterprise clients, while ByteDance pauses Seedance 2.0’s rollout over copyright disputes with major studios. That leaves headroom in video generation that Alibaba could exploit, building on its Qwen family and prior AI integrations across e-commerce, ads, and entertainment. ¹

The leaderboard win sharpens China’s competition in generative video and spotlights talent mobility: SCMP notes HappyHorse’s creator Zhang Di returns to Alibaba in November and leads the months-long effort to release the model still in internal beta. The benchmark’s design—dynamic ratings tuned by relative performance—makes gains visible quickly when human preferences tilt toward a newcomer. ²

HappyHorse tops global ranking after debut, per WSJ

After its debut earlier this month, HappyHorse 1.0 holds the No. 1 position on Artificial Analysis’ text-to-video leaderboard, signaling growing competitiveness from Chinese firms in media, advertising, and entertainment use cases that need high-quality generative video. WSJ frames the milestone against ByteDance’s earlier Seedance 2.0 release, underscoring the pace of iteration. ³

The leaderboard result builds on the same blind, head-to-head human preference tests that propelled the model’s rapid rise, matching CNBC and SCMP reports on its cross-category performance. The sustained No. 1 slot suggests quality that generalizes across prompts rather than a narrow benchmark quirk. ³

While Alibaba has shipped video-capable models before, WSJ notes none drew comparable attention this fast; the combination of an anonymous start and immediate top ranking created outsized buzz before the company stepped forward to claim authorship. ³

Open Source & Repos

MemPalace: A high-scoring AI memory system, free to use

MemPalace is a memory layer for AI assistants that promises to keep everything and make it findable, instead of letting the AI decide what to forget—an approach the project markets as “the highest-scoring AI memory system ever benchmarked.” It organizes content like a “palace” of wings, halls, and rooms; the repository is public and free to use. ⁴

Independent practitioners experimenting with local stacks describe real-world tradeoffs: a recent write-up evaluates MemPalace but ultimately opts for a simpler mem0 + Qdrant setup, focusing on debuggability and KV-cache tricks to stretch context on modest hardware. The post shares concrete settings for long context on a Mac mini (e.g., asymmetric KV-cache quantization in llama.cpp forks) to make retrieval-augmented chats practical. ⁵

Community proposals are shaping MemPalace’s roadmap in the open: one issue argues for hybrid search via Reciprocal Rank Fusion (RRF) and optional cross-encoder reranking, reporting measured gains on a production corpus (MRR 0.5395 → 0.8833; Hit@1 46.7% → 80.0%) using ChromaDB + SQLite FTS5. Another suggests an MCP tool to archive raw conversation excerpts verbatim—useful when the original reasoning is more valuable than a summary. These are issue discussions with shared code and measurements, not yet official releases. ^{6 7}

Research Papers

ClawVM: Virtual memory for stateful, tool-using agents

Many AI agents lose track of what happened earlier in a session or across sessions; ClawVM proposes a virtual memory layer that treats an agent’s context as typed pages with minimum-fidelity guarantees, multi-resolution representations under a token budget, and validated writeback at every lifecycle boundary. Because the agent harness already assembles prompts and mediates tools, the paper argues it is the natural place to enforce residency and durability as an auditable contract. ⁸

In tests spanning synthetic workloads, 12 real-session traces, and adversarial stress, ClawVM eliminates all policy-controllable faults whenever the minimum-fidelity set fits within the token budget (verified by an offline oracle), and adds a median under 50 microseconds of policy-engine overhead per turn—small enough to keep agents responsive. For readers, this means fewer “lost memory” failures and more predictable, cheaper runs. ⁸

Operational visibility complements memory policy: the companion ClawTrace plugin for OpenClaw records every run as a tree of spans, surfacing token burn, tool-call loops, and per-step inputs/outputs, with an AI analyst (“Tracy”) that queries the trace graph. Together, robust memory contracts and traceability help contain costs and failures in long-running agents. ^{9 10}

Community Pulse

Hacker News (67↑) — Interest in MemPalace’s “store everything” approach is mixed, with skepticism focused on the project’s benchmark claims after independent tests reported much lower end-to-end QA performance.

Why It Matters

High-quality video generation is consolidating around a few fast-improving contenders, and Alibaba’s confirmed authorship of a top-ranked model signals that leadership can emerge through results-first releases. That shapes which tools creatives and marketers may actually adopt next. ^{1 3}

At the same time, memory is becoming the new performance bottleneck for agents: systems like ClawVM and practical tracing tools are raising expectations for reliability, cost control, and auditability—key for putting agents to work on longer, more consequential tasks. ^{8 10}

Try This Week

1. Build your own memory-aware agent trace: install the ClawTrace OpenClaw plugin and inspect token usage and tool-call loops in your next run. https://github.com/richard-epsilla/clawtrace
2. Kick the tires on MemPalace: clone the repo and try organizing a small chat export to see how “wings/halls/rooms” retrieval feels on your data. https://github.com/MemPalace/mempalace