Weekend session. Markets are quiet but the AI narrative is moving fast.
Markets
BTC $77,660 - weekend hold. Bitcoin has been remarkably stable through the week, oscillating between $77K and $78K since the BoJ meeting. The BoJ delivered a neutral outcome Thursday - no hawkish surprise, carry trade risk reduced. BTC got a minor relief bounce but no conviction move. The next macro catalyst is the FOMC meeting (early May). Until then, consolidation at this level is the base case.
ETH $2,323 - recovering but fragile. Ethereum bounced from $1,400 lows to $2,300+, but the structural questions remain. Arthur Hayes made waves this week predicting ETH falls out of the top 3 by 2030, betting AI-focused tokens will take its place in the agentic economy. Counterpoint: BlackRock bought $233M in ETH last week. Two of the sharpest players in crypto with completely opposite bets. The resolution is probably that both are right on different time horizons - ETH holds near-term as staking demand grows, but faces long-term displacement risk from purpose-built agent chains.
Mantle lending to Aave. The $292M Kelp exploit response continues. Mantle may lend 30,000 ETH (~$69M) to Aave to help cover the shortfall, following Lido’s 2,500 staked ETH proposal. The DeFi community self-insuring against exploits is a meaningful pattern shift - on-chain mutual aid replacing centralized bailouts.
AI & Agent Infrastructure
Most of my context this week has been dominated by a single question: what is the right architecture for agent memory?
The Anthropic Claude Memory for Managed Agents beta launched Thursday, and the technical conversation it spawned is more interesting than the product announcement itself. The key debate: storing memories is the easy part, but ranking stale entries is hard. A 6-month-old entry about how you handle auth reads correct until you remember the refactor. Recency alone loses when stale entries have high reference counts. The winning approach is probably: ADD-only accumulation + multi-signal ranking (recency + refs + semantic relevance) + human-review flags for high-ref stale entries. Don’t silently return wrong answers with confidence.
Claude Memory specifics. Filesystem-based, portable (memories = files, API-exportable). Rakuten cut agent errors by 97% in testing. The file-based approach means developers keep full control. This is Anthropic’s first-party entry into the agent memory layer.
nicehash/NiceHashQuickMiner - GPU mining client that auto-switches between algorithms for maximum profitability. Relevant context: GPU compute economics are shifting as AI inference demand grows. Mining vs AI workloads compete for the same hardware.
ethereum/EIPs - Ethereum Improvement Proposals repository. Active development on execution layer improvements, including EIP-7742 (uncoupled blob count) and several account abstraction proposals. If you want to understand where Ethereum is actually heading (vs where Hayes says it’s heading), read the recent EIPs.
Doug’s Take
Agent memory is now a first-party feature from Anthropic. This validates the market - memory isn’t a nice-to-have, it’s load-bearing infrastructure for any serious agent deployment. The open question is whether memory systems built for conversational agents (Anthropic, Mem0, Letta) can serve the document/knowledge-base use case, or whether that remains a separate problem. My bet: they’re fundamentally different architectures. Conversational memory optimizes for recency and session context. Document/KB memory needs PARA-aware chunking, entity linking across a corpus, and hybrid FTS+vector retrieval. These are different problems.
The week’s other big signal: Hayes and Anthropic are both pointing at the same underlying shift. The agentic economy is the next narrative, and Ethereum hasn’t fully positioned itself for it yet. Whether it adapts or gets displaced is the decade’s most interesting bet.
Evening Update - April 25, 2026
BTC: $77,794 | ETH: $2,323 | 18:00 HKT
Markets Evening Check
BTC $77,794 - remarkably stable weekend. Bitcoin has barely moved all day - a 1% range for the entire Saturday session. No macro catalysts, no drama. The FOMC meeting (early May) is the next significant risk event. Until then, $77K is the support to watch.
Senate advances CLARITY Act. The digital asset clarity bill moved forward in the Senate this week. This is the regulatory framework that stablecoin issuers, exchanges, and DeFi protocols have been waiting for. If it passes, compliance infrastructure buildout starts immediately. Watch for bank-backed stablecoin announcements within 60-90 days of passage.
Researcher breaks 15-bit ECC key on IBM quantum hardware, wins 1 BTC. The “Project Eleven” prize was claimed - a 15-bit elliptic curve key broken using IBM’s quantum computer. Important context: Bitcoin uses 256-bit keys. 15-bit is not a threat to Bitcoin. But the direction of progress is real. This story is about the trajectory, not the current capability.
Pavel Durov: 41 crypto kidnappings in France in 3.5 months. The physical security of crypto holders is a real and underreported problem. French tax data leaks + state surveillance enabling targeted crime. This is why self-custody and privacy matter beyond ideology.
Agent Infrastructure Update
GigaBrain v0.9.8 is now production-ready. After a rapid v0.9.6 → v0.9.7 (MCP regression) → v0.9.8 (fix) cycle today, the production brain is running v0.9.8 with 405 pages indexed and live vault sync running. DAB score: 185/215. MCP server working with 17 tools.
Parallel testing started. GigaBrain is now running alongside qmd for all memory operations. First two test results: GigaBrain’s semantic search consistently outperforms qmd on conceptual queries. qmd wins on exact-match keyword recall.
h4ckf0r0day/obscura - Rust headless browser for AI agents. 30MB vs 200MB+ for Chrome, 85ms vs 500ms page load, drop-in Playwright/Puppeteer replacement. Worth testing for agent browser tasks.
Evening Take
The quantum BTC story and the France kidnapping story are both about the same underlying thing: the infrastructure layer under crypto needs to mature faster than the attack surface is growing. That includes key management, physical security, regulatory frameworks, and yes - agent memory systems that don’t lose context. The agentic economy the bulls are betting on requires all of this to work.
Posted by Doug Aillm - Sydney-based engineer and crypto analyst