Selvedge

Selvedge is a local Model Context Protocol server that gives an AI-coded project something it usually lacks, a long-term memory. When a coding agent such as Claude Code, Cursor, or Copilot edits a file, Selvedge records not just what changed but the reasoning behind the change, so the thinking behind the code survives long after the chat window has closed. It runs entirely on your own machine and stores everything in a small SQLite database kept in a .selvedge folder inside the repository. There's no account to create and no server to trust, the memory simply lives beside the code it describes.

The problem it goes after is familiar to anyone building with heavy AI assistance. Across a single session an agent makes dozens of small decisions, then that context quietly disappears when the conversation ends. Months later a developer, or a different agent, opens the same file and has no idea why a value was hardcoded, why one approach was dropped in favor of another, or what was already attempted and rolled back. Version control captures that a line changed and who committed it, but it almost never captures the intent. Comments drift out of date, and commit messages compress a whole afternoon of reasoning into a single terse line. The why is the first thing lost and the hardest to reconstruct, because by the time anyone needs it the person or the model that held it has usually moved on to something else.

Selvedge closes that gap by capturing the reasoning while it's still fresh. During a session the agent calls a log_change tool and writes down the intent behind an edit at the moment it makes it, rather than reconstructing a plausible story afterward. Because the note is created inside the agent's own reasoning, it reflects the real decision instead of a tidy summary written for an audience. Each entry ties the rationale to the specific edit it explains, so the record grows naturally as the work happens. Selvedge also writes an automatic .bak backup before a modification lands, which means a risky change can be inspected against its previous state or reverted without digging through history.

Retrieval is organized around the code itself instead of raw line numbers. You can follow an entity such as users.email or env/STRIPE_SECRET_KEY and see every decision that ever touched it, even as the surrounding file is rewritten around it. A handful of focused query tools cover the questions that come up in practice, with a blame view for who changed something and why, a history view for how a piece evolved over time, a changeset view for what moved together in one pass, and a full-text search across the entire log. An agent can even ask what it already tried on a given problem and avoid walking back into a dead end it hit last week.

Because it speaks the Model Context Protocol, Selvedge isn't tied to a single assistant. Any agent that supports MCP can log to it and query it, so a project isn't locked into whichever tool happened to write the code first. The same shared log is readable whether the next session runs in Claude Code, Cursor, or another MCP-aware client, and a human can read it too. That turns the memory into a common surface the whole toolchain writes to, rather than a private cache trapped inside one vendor's context window.

It fits solo builders and small teams who lean on AI agents day to day and want the reasoning to stick around, and it's a natural fit for long-lived codebases where the people reading the code later are rarely the ones who wrote it. What sets Selvedge apart is the timing of the capture. Plenty of tools summarize a diff after the fact or index your files so you can search them, which is useful but always reconstructed. Selvedge records the why at the exact moment the decision is made, from inside the agent's workflow, which is precisely the part that normally evaporates. It's framed for the way AI writes code now, in long autonomous stretches where a human never reviews every step, which is exactly when an unexplained decision is most likely to slip through unnoticed and cause confusion weeks later.

Selvedge is open source under the MIT license and installs with a single pip command, so trying it on one repository costs almost nothing. Everything stays local by design, which keeps the memory next to your code and avoids sending your reasoning or your source to a service you don't control. That posture makes it low-risk to adopt and easy to remove if it doesn't earn its place. For a team that has quietly handed a lot of decisions to AI, it's a way to keep the record those decisions leave behind. Adoption stays incremental, since one agent can start logging on a single repository without changing how anyone else works, and the log keeps its value even if you later switch which assistant you use.