Pig.dev, a participant in Y Combinator’s Winter 2025 batch, initially focused on building an AI agent capable of automating tasks on Microsoft Windows desktops. However, in May, founder Erik Dunteman announced a strategic pivot: discontinuing the “Pig” project and instead launching Muscle Mem, a caching system designed to help AI agents offload repetitive tasks for better efficiency and focus
In a candid post, Dunteman said his previous modela cloud API product failed to gather user interest, and a pivot to developer tooling didn’t attract demand either. Customers instead wanted bespoke automation consultants to build tools for them, not off-the-shelf development kits. That mismatch pushed Dunteman to rethink the product direction entirely.
Dunteman’s new focus, Muscle Mem, offers a caching layer for AI agents to store and reuse outputs from repetitive tasks. This allows agents to dedicate more resources toward fresh reasoning and edge-case processing. Dunteman emphasized that Muscle Mem remains inspired by the computer‑use problem, but tackles it from a developer tooling angle rather than trying to automate the entire OS directly
On a recent YC podcast, partner Tom Blomfield compared Pig.dev’s vision to the startup Browser Use, which helps agents navigate web interfaces. Both aimed to solve the “last mile” of human-computer interaction. However, building long-context agents for Windows desktops proved technically challenging: as usage time grows, reasoning accuracy drops, and LLM costs escalate, making sustained automation unreliable and expensive
Founders & Analyst Commentary:
Tom Blomfield advised founders to focus on narrowly defined enterprise problems rather than broad automation platforms. He sees value in vertical enterprise adoption over horizontal consumer targeting.
Erik Dunteman acknowledged that building Windows agents remains one of AI’s toughest challenges—but believes his caching approach may yet make strides in tackling long-duration task interpretation without the complexity of full automation.
