Andrej Karpathy posted something deceptively simple this week.

Dump raw research into a folder. Let an LLM organise it into a wiki. Ask questions. File the answers back in.

That's it. No RAG pipeline. No vector database. No fine-tuning. Just a folder, a model, and a loop.

But here's what makes it interesting: every query makes the wiki better. The knowledge compounds. You're not just retrieving information — you're building a second brain that builds itself.

Four independent signals this week converged on the same architectural bet. Karpathy's post on ephemeral multi-agent wikis (frontier models spawning research teams that dissolve after each task). @jumperz: "agents that own their knowledge layer don't need infinite context windows — they need good file organisation." @tom_doerr: obsidian-mind, the practical implementation with 896 likes. @DataChaz: "a living AI knowledge base that actually heals itself," 948 likes, 112 RTs.

When four unconnected people converge on the same architectural insight in the same week, it's not coincidence. It's a signal.

The implication for anyone building with agents: the expensive part isn't the model. It's the memory. And memory, it turns out, doesn't have to be expensive. A folder, a model, and a loop can outcompete a $50K RAG pipeline if the loop is designed well.

This is the architecture of Q2 2026. The teams building it quietly are 12 months ahead of the teams still debating vector databases.

Curated from the SignalMesh X intelligence sweep — Mar 28–Apr 3

[COVER SIGNAL] @karpathy — ephemeral multi-agent wikis

Karpathy described the natural extrapolation of LLM reasoning: frontier models spawning teams of LLMs to build ephemeral knowledge bases, then dissolving. He uses this pattern himself for research. The key insight: you don't need persistent compute infrastructure to maintain persistent knowledge. The wiki survives. The agents don't.

Why it matters: This pattern is cheaper, faster, and more inspectable than any RAG system. It's also how the most sophisticated agent deployments are being built right now.

[IMPORTANT] @tom_doerr / breferrari — obsidian-mind (896 likes)

Obsidian vault template for Claude Code. Start session → Claude reads your North Star, active projects, and recent memory. /dump auto-files meeting notes, decisions, and brag docs. Every session compounds the last. Karpathy's pattern made practical — on GitHub today.

github.com/breferrari/obsidian-mind

Why it matters: The gap between interesting idea and running in production just closed. This tool is the on-ramp.

[IMPORTANT] @quantscience_ — Ziplime

Modernised Zipline fork for algorithmic trading — updated Python/pandas compatibility, native AI strategy generation. Eliminates the friction between strategy ideation and execution in your terminal.

pypi.org/project/ziplime

Why it matters: The tooling gap between retail and institutional algo trading is closing. Ziplime is one of the tools doing the closing.

[IMPORTANT] @steipete — AI-generated security reports up 5x

AI-generated kernel security reports went from 2–3/week to 10+/week in two years. OSS maintainers are drowning in noise. At the same time, Emollick confirmed: prompt injection in CVs and letters of recommendation works on LLM judges. The same intelligence that makes AI useful makes it gameable.

Why it matters: Security and trust are now the core product problem in AI systems.

[IMPORTANT] @sharbel — voice AI and self-evolving agents

Weekly repo tracking surfaced a concentrated theme: microsoft/VibeVoice (+11.1K stars), bytedance/deer-flow (+9K), NousResearch/hermes-agent (+8.8K). Agents that remember, adapt, and act without being told twice.

Why it matters: The architectural direction is clear. Agents that persist and evolve are replacing agents that reset.

[TRADING] Polymarket — Will China have the #1 AI model by EOY?

New market, early liquidity, non-trivial implied probability. DeepSeek R2 rumours. Tariff-driven tech nationalism accelerating domestic AI investment. The market is pricing a real possibility.

Why it matters: If you have a view on the US-China AI race, this is a liquid place to express it.

Issue #007 baseline — star counts as of April 3. Weekly deltas begin Issue #008.

1. microsoft/VibeVoice — +11,100 stars this week

Clone any voice. 60-minute transcription. Fastest-growing AI repo this week by a significant margin.

2. bytedance/deer-flow — +9,000 stars this week

ByteDance's SuperAgent — researches, codes, and creates autonomously. Competing directly with OpenAI's operator-class products.

3. NousResearch/hermes-agent — +8,800 stars this week

Self-evolving memory agent. Rewrites its own memory architecture based on what it learns. Watch this one closely.

4. mvanhorn/last30days-skill — +8,600 stars this week

Research skill pulling signal from Reddit, X, YouTube, HN, and Polymarket simultaneously. The multi-source intelligence sweep in a single skill.

5. TauricResearch/TradingAgents — +3,900 stars this week

Multi-agent LLM trading framework. Distinct roles — analyst, risk manager, executor — coordinating on live market decisions.

6. google-research/timesfm — +2,800 stars this week

Zero-shot time-series forecasting. No fine-tuning required.

7. SakanaAI/AI-Scientist-v2 — +2,000 stars this week

Automated scientific discovery via agentic tree search. The research loop is becoming autonomous.

Three tools that shipped quietly this week. Each one solves the same problem: the gap between powerful AI primitives and production-ready workflows.

obsidian-mind — the knowledge layer that compounds

Every Claude Code session reads your North Star, active projects, and recent memory. /dump files decisions automatically. Your AI knows what you decided last Tuesday because it was there.

github.com/breferrari/obsidian-mind

Ziplime — backtesting meets AI strategy generation

Updated Zipline fork with Python 3.12 compatibility and native AI strategy generation. If you're running a backtester from 2018, this is the upgrade.

pypi.org/project/ziplime

openclaw-ops — production-grade agent operations

Full ops layer for OpenClaw: heal.sh, watchdog.sh, security-scan.sh, skill-audit.sh. 5-minute auto-heal that survives reboots.

github.com/cathrynlavery/openclaw-ops

Hyperliquid RWA hits $1.3B open interest — record

$1.3B OI, $1.4B weekend volume. When traditional markets close at 4pm Friday, Hyperliquid doesn't. This is a structural shift, not a volume spike. The 24/7 market for RWAs is being built in real-time.

$AZTEC perps go live on Hyperliquid at 3x leverage

New perpetuals listing. Hyperliquid continues expanding while competitors debate RWA roadmaps. The pace of listing expansion is itself the signal.

Microsoft commits $10B AI infrastructure to Japan

Third major Big Tech investment in non-US AI compute in 90 days. The inference compute war is becoming geographically distributed. Whoever owns the compute owns the margin.

OpenClaw 2026.4.2 ships Durable Task Flow orchestration

Native fleet orchestration — spawning and coordinating sub-agents without managing infrastructure manually. The same architecture Karpathy described. Not a coincidence.

The self-modifying agent is no longer theoretical

hermes-agent rewrites its own memory architecture based on accumulated experience. Not retrieving from a fixed store — reshaping the store. Agents that improve without retraining.

ByteDance's deer-flow enters the autonomous agent race

Research, code, create — competitive with the current frontier. The Polymarket "China #1 AI model by EOY" market is pricing this correctly.

AI-generated security reports overwhelm OSS maintainers

CVEs went from 2–3/week to 10+/week. Most are noise. Maintainers triaging AI output instead of fixing real bugs. Signal quality in security research is collapsing.

Prompt injection in professional documents works

Emollick confirmed: prompt injection in CVs, academic papers, and letters of recommendation manipulates LLM-based evaluation systems. As AI screens humans, humans learn to manipulate AI. The trust layer is the next infrastructure problem.

Will a Chinese company have the #1 AI model by EOY 2026?

New market. Non-trivial probability. DeepSeek R2 + deer-flow + tariff-driven domestic investment. If you have a view on the US-China AI race, this is where to express it.

Will there be no change in Fed interest rates?

Market consensus: no cut at next FOMC. Low-variance positioning for capital preservation.

Will Trump visit China by April 30?

Tail risk. Low probability, high news-flow sensitivity. Watching for tariff war escalation signals.

Prices as of Thursday April 3, 2026 | 21:00 CET

The Memory Efficiency Curve

The conventional assumption: more capable AI equals more expensive AI. The emerging data says otherwise.

obsidian-mind users report Claude Code sessions that know project context from session one — because the prior session filed the notes. The cost of that context: a /dump command and a markdown file. The alternative — keeping a 128K context window populated across every session — costs orders of magnitude more at API rates.

Karpathy's ephemeral wiki pattern compounds this: a fleet of small models with good file organisation outperforms a single large model with a giant context window on research tasks, at lower cost, with better inspectability.

The data point: the memory advantage is not a function of parameter count. It's a function of loop design.

The moat in 2026 is not model access — everyone has model access. It's the memory architecture running underneath.

The theme this week is memory. Not in the neuroscience sense — in the engineering sense. The agents that dominate the next 12 months won't have the largest context windows. They'll have the best loops.

Karpathy sees it. obsidian-mind ships it. hermes-agent evolves it. The repos growing fastest this week all point the same direction.

Build the loop. File the notes. Let it compound.

Next issue: the prediction market economy — how autonomous agents are becoming the most active participants in real-money forecasting markets, and what that means for signal quality.

Subscribe: signalmesh.beehiiv.com

Follow: @TheBlockchain

SignalMesh is produced with AI assistance. All analysis reflects editorial views. Not financial advice. Market data as of April 3, 2026 21:00 CET. Intelligence sweep: Mar 28–Apr 3, 2026.