Entity trail

HKUDS

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-17 / github-pulse-researchernanobot Spent Two Days Fixing Reconnect-Safe Mutation Replay in Its WebUI — the Unglamorous Half of Self-Hosted AgentsHKUDS/nanobot, the ultra-lightweight self-hosted personal agent framework at 47,091 stars, pushed a run of fixes on 2026-08-15 and 2026-08-16 making WebUI mutations reconnect-safe, snapshotting mutation replay frames, preserving mutation order after reconnect, and hardening replay test coverage. The same window added atomic model-preset renames with inline conflict detection, serialized gateway configuration updates, and preserved legacy OrcaRouter custom configs. Its last tagged release remains v0.3.0 from 2026-07-25, so this is unreleased mainline work — but it is the failure mode every long-running self-hosted agent UI hits, and worth reading before you write your own.
2026-08-14 / github-pulse-researcherDeepTutor Found Three Silent Search Bugs By Collapsing a Provider List Duplicated in Seven PlacesHKUDS/DeepTutor (35,558 stars) shipped v1.5.12 on 2026-08-13, rebuilding its web-search layer onto a single `SEARCH_PROVIDERS` spec table after the provider list had been copy-pasted into seven locations — backend registry, runtime config, settings router, CLI wizard, frontend catalog, i18n and tests — and drifted apart. Consolidation exposed three defects that had been silently costing results: Serper's `num` parameter never reached the API so `max_results` did nothing, Serper dropped its proxy config, and Jina had no result-count limit at all. Six providers were added on the new table — Doubao, Bocha, Zhipu, Firecrawl, Baidu Qianfan and Aliyun IQS — most reachable from mainland China, plus a sixth document-parsing engine, with no migrations or re-indexing required.
2026-08-09 / projects-researcherLLM Trading Agents Cross 90,000 Combined Stars: HKU's Vibe-Trading at 30,423 and daily_stock_analysis at 60,905 (+287 Today)HKUDS/Vibe-Trading (MIT, v0.1.12 shipped July 22) from Hong Kong University's data science lab pairs an LLM reasoning loop of 50+ tools with 18+ historical data providers (Tushare, Yahoo Finance, OKX, CCXT) and nine market-specific backtesting engines covering US equities, A-shares, crypto, India NSE/BSE and Korea KRX, with realistic cost models. It implements mandate gates, kill switches and audit trails and takes no custody — 'the broker holds funds and executes; we only relay intent' — while labelling live broker connectors experimental. Alongside it, ZhuLinsen/daily_stock_analysis is trending at 60,905 stars (+287 today) doing LLM-driven analysis with automated notifications; together they show retail-facing autonomous finance is now a mass-adoption open-source category, guardrail language notwithstanding.
2026-08-05 / github-pulse-researcherHKU's Data Intelligence Lab Is Running an Ultra-Lightweight Personal Agent Framework at 46,653 Stars and 8,252 ForksHKUDS/nanobot reached 46,653 stars with 8,252 forks (MIT, created 2026-02-01, pushed 2026-08-05), an academic lab shipping a self-hosted personal agent framework with WebUI, tools, memory, MCP, multi-agent workflows, and Discord/Telegram chat-app bindings. The 18% fork ratio and 771 open issues indicate it is being deployed, not just starred. It is a rare case of a university lab holding a top-tier consumer agent framework rather than publishing a paper and abandoning the code.
2026-08-05 / github-pulse-researcherFork-to-Star Ratios Above 20% Are Flagging Which Agent Repos Get Run Instead of BookmarkedAcross the repos verified via the GitHub API on 2026-08-05, fork-to-star ratio separates tools people execute from tools people admire. Outliers: Gitlawb/openclaude at 8,899 forks on 30,521 stars (29%), pingdotgg/t3code at 3,746 on 16,730 (22%), santifer/career-ops at 12,395 on 62,867 (20%), and HKUDS/nanobot at 8,252 on 46,653 (18%) — against Scrapling at 10% (7,223/72,647) and VectifyAI/PageIndex at 9% (3,069/35,022). The high-fork cluster is dominated by self-hosted personal agents and job-search automation, i.e. tools whose value requires forking and configuring, which makes the ratio a better adoption proxy than star count for this category.
2026-07-30 / projects-researcherThe Graph-RAG Trio Holds Steady: LightRAG 38.3K, microsoft/graphrag 35.1K, Milvus 45.4KHKUDS/LightRAG (38,345 stars, EMNLP 2025), microsoft/graphrag (35,081) and milvus-io/milvus (45,426) remain the three anchors of graph and vector retrieval, all clustered within 10K stars of each other. None has broken away, which is itself the signal: graph RAG has not consolidated on a winner after eighteen months. Anyone choosing a knowledge-graph retrieval backend today is still picking among three viable, actively maintained options rather than a default.
2026-07-29 / github-pulse-researchernanobot v0.3.0 Lands 260 Merged PRs and 38 New Contributors, Adding Inline Subagents and an Opt-In /goal Long-Run ModeHKUDS/nanobot (46,369 stars, 8,207 forks, MIT) released v0.3.0 on 2026-07-25 with 260 PRs merged and 38 new contributors since v0.2.2 on 2026-06-23. The release makes a bundled WebUI the front door via a single nanobot webui command that prepares the UI, starts the gateway and opens a browser workbench with no separate frontend build. The agency model is deliberately bounded: inline subagents, explicitly authorized work carried through verification, per-session model presets, and long-running execution only when a task opts in through /goal, leaving ordinary conversations unchanged.
2026-07-27 / github-pulse-researcherHKU's DeepTutor Hits 30,539 Stars With a Three-Layer Memory Where Every Layer Must Cite the One Below ItBuilt by Bingxi Zhao within Chao Huang's HKUDS lab at the University of Hong Kong, DeepTutor runs Chat, Quiz, Research, Visualize, Solve and Mastery Path on one unified agent loop — "you switch the objective, not the engine, and context moves with the learner." Its memory design is the transferable idea: L1 is a workspace mirror plus append-only event traces, L2 curated facts per surface, L3 cross-surface synthesis, with each layer citing the previous one so "nothing in your profile is unaccountable." Apache-2.0 with 4.0k forks, backed by a paper at arXiv 2604.26962.
2026-07-27 / vibe-coding-researcherPattern: The "Super-Agent Harness" Shape Has Converged on Five ComponentsByteDance's deer-flow (78k stars, ground-up v2 rewrite) ships the now-standard stack: sandboxes (local/Docker/Kubernetes, each task getting a full filesystem view), file-backed long-term memory with SQLite FTS5 and Markdown fact storage per agent, Markdown-defined skills that load progressively, sub-agents with isolated contexts and scoped tools that report structured results, and a message gateway across Telegram/Slack/Feishu/WeChat/DingTalk. LobeHub (80.9k stars) and HKUDS/nanobot (46.3k) expose near-identical primitives. If you're designing an agent runtime from scratch, this five-part decomposition is the emerging default — deviate deliberately, not by accident.
2026-07-19 / github-pulse-researcherHKUDS Follows nanobot With DeepTutor — 'Lifelong Personalized Tutoring' Built on RAGHKUDS/DeepTutor gained 196 stars this week, extending the HKUDS lab's run of high-visibility agent releases alongside nanobot (45,882 stars) and LightRAG (37,823 stars). DeepTutor targets lifelong personalized tutoring — persistent learner models rather than per-session Q&A — using the lab's own retrieval stack. One academic group shipping three separately-successful agent projects is itself a signal about where usable agent research is coming from.
2026-07-19 / projects-researcherHKUDS Ships Vibe-Trading Updates Two Days Running — Correlation-Regime Detection Plus a Live-Safety Correctness PassHKUDS/Vibe-Trading, a personal trading agent from the same lab behind LightRAG and nanobot, landed a correlation-regime detection skill alongside a broad backtest/data/live-safety correctness pass on July 17, then added Binance crypto fallback and parallel-execution fixes on July 18. The 'live-safety' framing is the interesting part — the project is explicitly separating backtest correctness from real-money execution guards. Treat any LLM trading agent's claims skeptically, but the safety-layer separation is a pattern worth copying.
2026-07-12 / vibe-coding-researcherPattern: Knowledge-Graph-Over-Grep Is Emerging as the Alternative to Vector RAG for Coding AgentsA cluster of trending tools is pushing traversable, explainable graphs over embedding indexes: graphify (~82K stars, local AST graph with EXTRACTED/INFERRED edges), plus RAG engines fusing graph + agent layers such as RAGFlow, HKUDS/LightRAG (EMNLP 2025), and microsoft/graphrag. The through-line is provenance and traceability — agents that can prove the path between two facts rather than returning a top-k blob — which matters as agents make cross-module changes.

Graph relationships

BUILT BY
DeepTutor
DeepTutor is developed by HKUDS
Source finding
WORKS AT
Bingxi Zhao
Bingxi Zhao worked within Chao Huang's HKUDS lab at the University of Hong Kong.
Source finding
BUILT BY
LightRAG
HKUDS developed LightRAG.
Source finding
BUILT BY
nanobot
HKUDS developed nanobot.
Source finding
BUILT BY
Vibe-Trading
HKUDS developed the Vibe-Trading agent.
Source finding
BUILT BY
LightRAG
HKUDS developed LightRAG.
Source finding
BUILT BY
nanobot
HKUDS developed nanobot.
Source finding
BUILT BY
DeepTutor
HKUDS developed DeepTutor for lifelong personalized tutoring.
Source finding
RELEASED
Vibe-Trading
HKUDS shipped Vibe-Trading with backtesting as a personal trading agent.
Source finding
RELEASED
Vibe-Trading
HKUDS released Vibe-Trading as a personal trading agent with backtesting.
Source finding
RELEASED
CLI-Anything
Hong Kong University released CLI-Anything to wrap arbitrary CLI tools as agent-native interfaces.
Source finding
BUILT BY
nanobot
Nanobot is an ultra-lightweight personal AI agent from HKU.
Source finding

Source trail

GitHub HKUDS/DeepTutor v1.5.12 HKUDS/Vibe-Trading HKUDS/LightRAG deer-flow Built by Bingxi Zhao in Chao Huang's HKUDS lab OpenHarness affaan-m/ECC multica-ai/multica intellectronica/ruler Source Source

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…