Priyank Patel

I kept switching between Claude Code, Cursor and Codex, and every new session asked me the same questions about the repo. Anchor is my fix. It’s a small MCP server that keeps one memory per project in a SQLite file on your machine, so the next agent starts with the facts and decisions instead of from zero.

Across eight test scenarios it cut 98,200 tokens of old transcript down to 2,239 and still kept every fact that mattered. It isn’t perfect: about one in six memories it pulls back wasn’t needed.

I wanted to know if the Recursive Language Models paper holds up on hardware like mine, a laptop with a 6 GB GPU, so I rebuilt it with a 2B Qwen model. The long document never goes into the prompt. It sits in a Python REPL, and the model writes code to search it and calls itself on the parts it needs.

At 64K tokens it found the needle 67% of the time, against 33% when the same model read the document directly. Up to 16K the plain model does as well or better, and a recursive answer takes about a minute instead of 15 seconds. All of it is in the write-up.

Give it a disease and it does the first pass of a drug-discovery search for you. It picks a target, pulls the protein structure, and looks for drugs that are already approved for something else but bind that target. Under the hood it’s a Gemini function-calling loop over UniProt, OpenTargets, RCSB PDB, AlphaFold DB and ChEMBL.

On six benchmark diseases it found the textbook target every time, in about 38 seconds each, without being handed any answers. It’s a way to narrow things down for a researcher, not a replacement for the lab.

Eight agent skills for reviewing pull requests the way a careful maintainer would. Most of each skill is about reading the change properly before saying anything. Every finding gets a score, the weak ones are dropped, and a review never leaves more than five comments.

npx skills add priyank766/OpenSource-SKILL

An agent that works inside a real project folder instead of a chat box. It can only touch files in the folder you give it, shell commands run in a Docker sandbox after you approve them, and you can watch every step it takes.