Side‑project surface narrows to memory

A lot of weekend AI side projects now look less like “build a chatbot” and more like “build a notebook the agent can actually remember.” Garry Tan’s new open-source project, GBrain, is pitched as a personal knowledge brain that imports Markdown files into Postgres and pgvector so an agent can search past notes instead of starting cold every session. (github.com) That shift showed up fast in the repo itself. GBrain was published about a week ago, had 3.7 thousand GitHub stars and 416 forks when crawled today, and its recent release notes mention incremental sync, file storage, and an installable skill instead of a broad consumer chat interface. (github.com) The basic problem is simple: a general chat window forgets everything unless you paste it back in. A memory layer works more like a filing cabinet with search, where notes from old meetings, plans, and documents can be pulled back into the next task. (github.com) The second signal is coming from coding tools that are narrowing the agent’s job on purpose. The Superpowers project describes itself as a complete software development workflow built from composable skills, and its README says the agent starts by pulling out a spec instead of immediately writing code. (github.com) That repo is moving at a very different scale from a hobby prompt pack. Superpowers had 143 thousand GitHub stars and 12.3 thousand forks when crawled today, and its recent updates include support for Claude Code, GitHub Copilot command line interface, Cursor, Gemini, and OpenCode. (github.com) A skill framework is basically a toolbox with labeled drawers. Instead of asking one giant assistant to do everything, developers give the agent a small set of repeatable moves like write a plan, edit one file, run tests, or install a capability. (github.com) (aitoolly.com) That is why memory and workflow are showing up together. Memory tells the agent what happened before, and a constrained workflow tells it what it is allowed to do next, which is much easier to inspect than a free-form chat log full of improvised steps. (github.com 1) (github.com 2) The practical side-project angle is narrower than it sounds. A founder can build an agent that remembers customer calls, a recruiter can build one that tracks candidate histories, and a developer can build one that follows a fixed plan-test-edit loop, all without pretending the product is a universal assistant. (github.com 1) (github.com 2) You can see the design preference in the details both projects emphasize. GBrain talks about searchable files, hybrid search, and a database-backed brain, while Superpowers talks about specs, chunks short enough to review, and modular skills that can be combined for one coding job at a time. (github.com 1) (github.com 2) That usually leads to plainer products. Instead of a glossy chat app with a thousand possible prompts, the winning demo in April 2026 is often a small agent with a memory store, a visible checklist, and a narrow promise like “remember my research” or “ship this feature safely.” (github.com) (aitoolly.com)