Blume turns repeated coding-agent corrections into persistent rules

What Blume does

A developer named Peder has released Blume, a free desktop application that watches coding-agent sessions on your machine and converts the corrections you keep repeating into durable rules, hooks, and skills. According to the launch post on dev.to, the app works with Claude Code, Codex, and Cursor, runs on macOS, Linux, and Windows, and processes everything locally — no code or chat transcripts leave the machine. It launched on Product Hunt this week.

The post frames the problem as "agent drift": an agent in a fresh session forgets conventions a human already explained, so the same corrections about architecture patterns or stray debug logging get typed over and over. The conventional fix is hand-maintaining CLAUDE.md, AGENTS.md, or Cursor rules, but the author argues that discipline decays within weeks. The post cites advice from Claude Code creator Boris Cherny — when the agent makes a mistake, have it write the fix into its context file — and notes that almost nobody keeps this up consistently. The naive alternative, having the agent append a rule after every correction, produces an ever-growing file of contradictory one-off instructions that makes the agent slower, more expensive to run, and less likely to follow its own context.

How it decides what becomes a rule

Blume reads the session histories that agents already store on disk and looks for three kinds of signal: direct corrections that reverse something the agent did, frustration expressed through short and escalating messages or unusually high token spend on a simple task, and softer recurring nudges such as reminders to run the tests. It clusters these signals by theme, and when a cluster crosses a threshold it proposes a concrete change — a rule for the context file, a hook that runs automatically, or a skill that packages a workflow the user keeps re-explaining.

The thresholds are deliberately conservative: five occurrences in the same thematic cluster, or two occurrences classified as pain, meaning they carried frustration signals or abnormal token usage. Anything below that stays an observation rather than a suggestion. Answering Product Hunt questions about one-off feedback hardening into permanent rules too early, the author says the system does not judge any single message — it waits for recurrence. A single frustrating session is just one data point; the same frustration appearing three times earns a rule.

Every proposal requires explicit human approval, and nothing is written to configuration without it. Each suggestion shows the cluster of session moments behind it, so the user can inspect the evidence before approving, editing, or dismissing.

What it does day to day

The main screen lists every agent session on the machine and its state — working, finished, or awaiting approval — so several sessions can run without watching terminals. Behind that, the app maps the files that shape agent behaviour across projects: rules, skills, hooks, and MCP servers. The author says many early users did not realise how many of these files they had, or which ones were still being read. Blume also tracks Claude Code and Codex plan usage against weekly limits, though the author calls that the least interesting feature. An analytics view showing whether a setup is improving over time is still in progress.

The post is candid about limitations. Blume is a desktop app rather than a CLI, so it is not scriptable for CI pipelines. It needs a week or two of real session history before it has anything useful to say, making it useless on brand-new projects. And solo developers who already maintain a tight, hand-tuned context file may not need it, since the problem grows with the number of sessions, contributors, and agents involved in a codebase.

Why it matters

Context files have quietly become the configuration layer of AI-assisted development, and they fail in two familiar ways: manual maintenance goes stale, and naive automation turns into unbounded noise. Blume's bet is that the answer is already in the data — session histories reveal which corrections people actually care about, because humans repeat what matters to them. Mining that pattern, applying conservative thresholds, and keeping a human in the approval loop is a design that could plausibly extend beyond the three supported agents. The local-only processing is also significant, since session logs contain proprietary code and unfinished ideas that should never leave a machine. It is early days and single-sourced — this is a developer's launch post, not an independent review — but the tool targets one of the most common complaints in agent workflows: re-teaching every session what the last one already learned.