Cardumen lives in your repo: memory, skills and verification installed once. Your next agent starts where the last one stopped.
Cardumen lowers the AI bill and raises consistency between sessions: five agents share one source of truth that lives on your machine.
Operational pain
On Tuesday you open a new session and your agent does not remember Friday. It rereads 18 files, repeats four prompts, tries the same idea again and runs out of context halfway through the fix. Claude Max cost $200 this month and twice left you halfway through a release. The side project has rules that live in your head, and every new chat pays for them again.
Cardumen response
Cardumen installs the context layer inside the repo: operational memory, versioned skills and a verify loop that does not let anyone declare done without evidence. The agent starts each session with Friday's decisions, not zero. Your cheaper model performs like an expensive one because the scaffold explains more variance than changing models.
Your next agent starts with Friday's context, not from zero.
−43% tokens in the first comparable task, measured.
Five agents (Claude, Codex, OpenCode, Cursor, Copilot) reading the same source of truth.
Closures with evidence, not convincing answers without proof.
Local-first: your code does not leave your machine to be useful.
Fit signals
You pay for two or more AI agents and want to consolidate without losing capability.
Your Claude Max quota or Cursor plan cut you off halfway through a release.
You have a serious side project and hate re-explaining it in every new chat.
You want to keep your code and decisions outside any cloud.
Once, inside the repo. You do not change editor, assistant or workflow.
Rules, gotchas and conventions stop living in chats and start living next to the code.
Every task closes with verify: tests, build or an explicit criterion. No self-deception.
Cardumen installs in your repo this week. Your next agent will already know what you decided on Monday.