Blume 1.0.56 turns the corrections you keep repeating into skills your coding agents actually follow.
There is a small, specific frustration that shows up once you start coding with agents every day. You correct one. "Do not reach for as any, validate it with zod." It fixes the code, thanks you, and moves on. Two days later, in a different repo, in a different conversation, it does the exact same thing. So you type the exact same sentence. Again.
The first time, it is fine. The tenth time, you start to feel like you are the one with the memory problem.
You are not. The agent is not broken either. The problem is structural, and it is worth looking at closely, because it turns out to be the whole game.
Where your instructions actually go
A coding agent's real memory is not the chat window. It is the rulebook: the CLAUDE.md at the root of your repo, the skills you have written, the hooks that fire on save, the scattered instructions files that tell the agent how this project wants to be treated. That is the part that persists. That is the part the agent reads at the start of every session.
Everything else, every correction you type into a conversation, is written on water. It steers the agent for that session and then it is gone. The knowledge was real. You paid for it with your attention. But it never made it into the one place that would have kept it.
So the same corrections come back. Not because anyone is failing, but because there is no path from the thing you keep saying to the rule the project keeps. The path exists in theory. You could stop, open CLAUDE.md, and write the rule down by hand. But you are mid-task, the agent is waiting, and editing your own config feels like flossing: obviously good, easily deferred. So the rulebook drifts out of date, and your corrections stay trapped in transcripts nobody will ever reread.
The corrections are the rules
Here is the part that took us a while to see clearly. The stuff you keep repeating is not noise on top of the work. It is the missing rulebook, dictated one frustrated sentence at a time. Every "no, use the existing helper," every "we do not do it that way here," every third reminder about the same import style is a project rule you already wrote. It is just stranded in a conversation instead of committed to a file.
If the corrections are already the rules, then the job is not to generate guidance from nothing. It is much more modest, and much more trustworthy: notice what you already said, often enough that it clearly matters, and offer to write it down properly. You stay the author. The tool just does the part you kept deferring.
That is Blume.
What Blume does
Blume is a small desktop app that sits alongside your coding agents (Claude Code, Codex, Cursor, Copilot, and the other supported harnesses) and does two quiet jobs.
First, it lets you see what every agent is doing. Which ones are running, which are waiting on you, which finished, and what rules and skills each one is operating under. If you have ever had four terminal tabs open and lost track of which agent needs an answer, this alone is worth the install.
Second, and this is the release we are most excited about, Blume turns your repeated corrections into proposed skills. It reads through your recent conversations, finds the correction, steering, and frustration signals that cluster into a pattern, and when the evidence is strong enough, it drafts a durable change: a skill, a line for CLAUDE.md, a hook, a doc. Then it shows you the proposal next to the actual messages it came from, and asks.
Not "here is what I changed." Instead, "here is what I noticed, want to make it a skill?" You read the evidence. You see the exact diff. You click Apply, Snooze, or Dismiss. Nothing touches your setup until you say so.
We think that distinction, proposal not action, evidence not assertion, is the difference between a tool you trust with your codebase and one you quietly turn off.
Improvements is a beta. Not every conversation becomes a suggestion, and Blume never edits your setup silently. The source messages stay visible the whole way through, and you make the final call.
What is new in 1.0.56
This release is mostly about making that loop, notice then propose then decide, fast and trustworthy enough to live with.
Improve got rebuilt to work incrementally. Instead of a heavy re-run over everything, Blume now processes new conversations as they arrive, groups related signals into coherent clusters, and keeps the source evidence attached from raw message all the way to finished suggestion. You can see why a proposal exists, not just what it says.
Global Search is new. One shortcut searches across conversations, messages, projects, setup artifacts, and suggestions. Keyboard filters narrow it down, and enter lands you back on the exact source with the match highlighted. It is the fastest way to answer "where did I decide that?"
Setup and Artifacts is far more complete. Project scans are faster, the inventory covers more supported artifact types, and each one gets a type-aware detail page with relationships, diagnostics, and history, so you can finally see what instructions, skills, and tools a project actually has, instead of guessing.
Agents got more dependable. Stopped sessions now resolve correctly instead of lingering as active. Running, idle, error, approval, and todo states are more accurate, Codex todo progress and Claude approval waits are represented honestly, ordering is steadier, and the whole panel flickers less.
First run is clearer, with provider-aware onboarding and better authentication guidance, so the first ten minutes feel like the rest of the app.
The principles under it
Your history stays on your device. Blume is local-first. Your conversations do not need to leave your machine for it to work.
You are the editor, not the edited. Blume does not run your agents, does not orchestrate them, and does not learn some private model of you in the background. It observes the work you are already doing and hands you a decision. The agency stays yours.
Effortless is a feature. Powerful developer tools have a bad habit of asking you to become an administrator of the tool itself. The whole point of Blume is to remove bookkeeping, not add a new console to babysit.
Why this matters more each month
We keep hearing the phrase "agentic-native codebase," and under the jargon there is something real. As more of your code gets written by agents, the rulebook that steers them stops being config and starts being one of the most important artifacts in the repo. It is the difference between agents that compound, each one a little more on-context than the last, and agents that keep relearning the same lessons and making the same mistakes forever.
Right now that rulebook is maintained by hand, badly, by people who are too busy to maintain it. Blume's bet is that it should not be. The knowledge to keep it current already exists. It is in the things you say every day. It just needs a path from your keyboard to the file. The point of all of it is simple: give every coding agent the same context.
Download or update Blume, then open Improve. It is free, it is local-first, and it runs on macOS, Windows, and Linux.
Get Blume at www.blume.codes
Blume is built by Peder Aaby and Olav Ljosland. We read everything. Tell us what you keep repeating to your agents, and what Blume should notice next.