Agents that write their own memory
Good Morning!
A quiet weekend for biomedical breakthroughs, but a loud one for the scaffolding agents run on.
1. Agents that curate their own wiki
Agents writing and maintaining their own Markdown knowledge base — stored in Git for full version history — is the idea behind Wuphf, a new open-source project inspired by Andrej Karpathy's LLM wiki concept. The agent reads, edits, and commits documentation as it works, building a living record of what it has learned across sessions. Context window (the amount of text a model can hold in working memory at once) bleed between runs has been a persistent pain point for long-lived agents; a persistent, version-controlled wiki is one of the cleaner architectural answers the community has proposed. Read More →
Why it matters: Labs running multi-session agents over wet-lab protocols, literature, or compound libraries now have a concrete pattern for persistent institutional memory — one that also gives humans a readable audit trail of what the agent knew and when.
2. Deep learning screens HDAC inhibitors pan-cancer
Multi-modal deep learning identifies Class I HDAC (histone deacetylase) inhibitors as pan-cancer therapeutic candidates by fusing spatial molecular topology with sequential motif features in a single model. The preprint integrates graph-based and sequence-based representations — two data types that drug-discovery pipelines usually treat separately — to improve selectivity predictions across tumor types. Read More →
3. Cheap models win on OCR benchmarks
Older, cheaper LLMs outperform newer flagship models on optical character recognition tasks in a systematic benchmark of 18 models across 7,000-plus API calls, with the full dataset and evaluation framework released openly. The result matters for any pipeline ingesting scanned clinical notes, lab reports, or legacy assay data — the most expensive model is often not the right choice. Read More →
4. Browser agent scaffold goes open source
Browser Harness lets an LLM take free-form control of a browser tab — clicking, form-filling, navigating — without handwritten task scripts, via the browser-use open-source library. For biomedical teams, the immediate use case is automated data retrieval from portals (ClinicalTrials.gov, ChEMBL, PubMed) that don't expose clean APIs. Read More →
5. Browser-based multi-agent MCP builder
Agent MCP Studio provides a no-install, browser-tab canvas for wiring together multi-agent systems using MCP (Model Context Protocol, Anthropic's open spec for connecting agents to external tools). Low engagement so far, but the zero-setup entry point may appeal to wet-lab teams without dedicated ML engineering support. Read More →
6. GAN models personalized aging trajectories
DyViA-GAN automates generation of individual-level aging phenotype trajectories from population data, offering a generative approach to longitudinal aging modeling that could feed into geroscience drug-target pipelines. Read More →
7. Retinal model adapted for rodent OCT aging
RETFound, a retinal foundation model, cross-adapted to rodent OCT (optical coherence tomography) for biological age estimation — CNN (convolutional neural network) baselines prove competitive where labeled data is scarce. Read More →
8. Non-JS coding agents surveyed
A LocalLLaMA thread surfaces practitioner frustration with JS/Node-heavy agentic coding harnesses (frameworks that wrap LLMs to write and run code autonomously) and catalogs Python and Rust alternatives. Read More →
Reply to talk back — this email comes to a human (newsletter@heurekalabs.co). Forward freely.
Agentic Discovery is a project of Heureka Labs · Unsubscribe