Git-based LLM wikis move agent memory into Markdown

TLDR: Git-backed LLM wikis, GPT-5.5 rollout, and new browser harnesses are all quietly reshaping how your agents remember, browse, and ship work.

Karpathy-style git wiki gives agents durable long-term memory

Wuphf is a new open source "LLM-native wiki" that stores agent knowledge in local Markdown files backed by Git, with a Bleve BM25 plus SQLite index on top. There is no vector or graph database yet, and the whole thing lives under ~/.wuphf/wiki/ so you can git clone your knowledge base out as needed.

This is aimed at the pattern Andrei Karpathy has been talking about: a persistent substrate that agents both read from and write into so context compounds across sessions instead of being re-pasted every morning. For builders frustrated with opaque RAG setups, the appeal is obvious: local, auditable text files, standard Git tooling, and a classical search index you can debug.

If you are experimenting with autonomous or semi-autonomous agents, this is a concrete way to prototype durable memory today, as of 2026-04-25, without committing to heavier infra.

Read more →

---

Browser Harness drops frameworks to give LLMs raw browser control

Browser Harness is a new open source project that wraps Chrome through the Chrome DevTools Protocol (CDP) and strips away most of the usual framework abstractions so large language models can operate the browser with maximal freedom. It builds on the existing Browser Use library, which already has tens of thousands of lines of deterministic heuristics for element extraction, click helpers, target management, watchdogs, and crash handling.

Instead of a rigid high level API, Browser Harness focuses on giving the model direct, low level tools plus the ability to self correct and add new tools that it is pre trained on. For anyone building serious web automation or research agents, this may reduce the amount of brittle glue code you write to translate between your agent and the DOM.

The flip side is that more freedom also means more room for the model to get lost, so expect to spend time tuning prompts and constraints in real deployments.

Read more →

---

OCTCube trains a 3D multimodal foundation model for eye imaging

A Nature paper introduces OCTCube, a three dimensional multimodal foundation model for optical coherence tomography (OCT) and related retinal imaging, trained on 26,605 3D OCT volumes that include around 1.62 million 2D OCT slices. The authors present three variants: OCTCube for uni modal OCT, OCTCube IR for bi modal OCT plus infrared retinal imaging, and OCTCube EF for tri modal input including fundus autofluorescence imaging.

On eight retinal disease prediction tasks, OCTCube achieved state of the art performance and showed strong generalization across cohorts, imaging devices, and modifiers that typically trip up narrower models. If you work on medical AI, this is another data point that domain specific foundation models can beat generic vision models once you have enough high quality data.

Regulatory clearance and real world validation will take time, but the paper offers useful architectural and training details for anyone building domain tuned 3D models as of 2026-04-25.

Read more →

Quick Hits

• OpenAI Unveils GPT-5.5 to Field Tasks With Limited Instructions GPT-5.5 targets more complex work, including operating email, spreadsheets, and calendars directly for scientists and developers. Treat this as a signal that computer-use agents are now a first class product surface.

• Google's AI Agentic Eyed By EU Watchdog 04/24/2026 The European Union is scrutinizing Google Cloud customer experience agents deployed at companies like Home Depot, Best Buy, and Papa Johns Pizza. If you sell agentic CX into Europe, expect more compliance and audit requirements.

• Top 10 uses for Codex at work OpenAI outlines concrete Codex workflows across tools, files, and business processes, which doubles as a decent idea list for internal agent pilots.

• GPT 5.5 on AI Gateway Vercel AI Gateway now exposes GPT-5.5 and GPT-5.5 Pro, tuned for long running agentic coding, computer use, and research, with better token efficiency for large systems.

• GPT-5.5 prompting guide Simon Willison summarizes OpenAI’s official tips, including a pattern where the model does hidden multi step tool planning before returning user visible results, which is highly relevant if you are orchestrating complex tools.

• Show HN: We built a way for Claude Code to join meetings like a real teammate Agentcall.dev lets existing coding agent sessions join Zoom, Google Meet, or Teams with voice, screen share, and live coding, closing the gap between "chat in a terminal" and actual pair programming with a team.

• llm 0.31 The llm CLI adds support for GPT-5.5 plus options to control verbosity and image detail, which is handy if your workflows still live in terminals or scripts.

• Working with Codex OpenAI’s guide walks through setting up Codex workspaces, threads, and projects so you can treat Codex as a persistent teammate instead of a one off chatbot.

• Codex settings A companion doc detailing personalization, detail, and permission settings, which matters if you are standardizing Codex across a team.

• Building Workforce AI Agents with Visier and Amazon Quick AWS shows how to connect the Visier Workforce AI platform into Amazon Quick via Model Context Protocol (MCP) so HR and ops staff can query workforce data through a single agentic workspace.

• DeepSeek V4 - almost on the frontier, a fraction of the price DeepSeek V4 Pro and V4 Flash launch as million token context models positioned close to frontier quality at lower cost, which is relevant if you are price sensitive on long-context workloads.

• Amazon Quick for marketing: From scattered data to strategic action Amazon Quick pitches a personal knowledge graph that unifies your marketing tools and data, effectively packaging an agentic layer around campaign analytics and planning.