ReplayAutonomous app testing for teams who ship at AI-speed
Promoted
Launched this week
Blume watches your coding agent sessions locally and turns what it learns into better agent context. Repeated corrections become rules, workflows become skills, and your agents stop making the same mistakes. Claude code, Codex and Cursor.
Blume.codes
Hey Hunters, Peder here👋
We built Blume after our previous startup suffered severely from agent drift. Duplicated functions, incoherent architecture and sneaky production bugs.
We tried to control it with all the best practices: proper rules, skills, docs, self-verification. But when our repo grew, the setup rotted fast. Manual maintenance is too time consuming and asking the agent to maintain it lead to context drift and bloat.
Believing the true context signals lie in human intent, we had the idea to extract human intention and corrections from sessions and use them to improve the repo context. After initial testing, we were sure we were onto something, so we pivoted the entire company and built Blume.
Blume is a desktop app that sits next to Claude Code, Codex and Cursor and learns from your sessions. Correction, nudges, frustrated ALL CAPS messages: Blume picks up on them and clusters them thematically. If a pain threshold or recurrence is reached, it suggests a concrete update to your setup. That can be updating a stale rule, patching a hook or skill, or creating an entirely new one. You review and apply it, improving the agent for the next run.
It is free to use, and uses your local harness for extracting and improving the setup. So no chats or code ever leaves your machine.
Try it now for macOS, Linux or Windows! And happy to answer any questions today on how it works or discussing approaches to giving agents proper up-to-date context
1752vc Pitch Deck Analyzer
This is exactly the failure mode I've hit with my own skill setups - one off feedback getting hardened into a permanent rule too early. How many repeats does Blume wait for before it turns something into a rule?
Blume.codes
@tmaleh_ It is quite conservative right now, either 5 occurences in a cluster (thematically grouped, so not necessarily the exact same problem). Or 2 occurrences of something that have been identified to cause "pain" (high token usage, all caps, frustration).
frustrated ALL CAPS messages getting clustered into a rule, ok that made me laugh. how do u separate a keeper rule from me just being annoyed at one file once
Blume.codes
@niveditha_patluri1 The product extracts signals from the chat, clusters them, and if any cluster reaches a pain- or recurrence threshold, then it is sent out to update your setup :)
Funny enough ALL CAPS RANT would actually influence the algorithm to give a high pain score / likely promotion to a change
Blume.codes
@niveditha_patluri1 I can also mention that processing happens with your harness, with the algorithms for promotion etc. being in the blume sidecar
Stop making the same mistakes is exactly what I want from agent memory. Very cool concept.
Blume.codes
@owen_carter3 That is the goal! We will work a whole lot on improvements going forward
This solves a problem I think a lot of solo builders are starting to run into. AI can move incredibly fast, but maintaining context, across dozens of sessions become its own job.
Blume.codes
@oscar_bennett Amen! When the repo grows you will experience gradual decrease in efficiency, which brings the initial fast momentum to a stall