Use cases / Engineering
Your agents start from what the team already decided
Coding agents forget everything between sessions. Korium gives them the decisions, the reasons and the gotchas, plus a map of the code, so every change starts from what the last one learned. Masthead, our open source plugin, runs this loop on Claude Code.
We build Korium this way. Every change searches memory before its first edit and saves what it learned before it stops. Nothing merges on a decision someone half remembers.
What goes wrong today
The agent rediscovers the codebase
Every session reads file after file to find what the last session already found.
The decision lived in a chat
Why the code works this way was settled in a thread nobody can find, so the next agent undoes it.
Review checks against a memory
The reviewer remembers the decision roughly, and roughly is how it gets enforced.
The same gotcha bites twice
One session lost an hour to it. The next session loses the same hour.
Before the change
The map says where. The memory says why.
The first thing Korium asks is what the team already knows about this part of the code. The agent gets one working brief, then reads the code itself before it changes anything.
After the change
The next session starts where this one stopped
What was decided, what bit, and what is left go into Korium before the session ends. The code map is refreshed after each commit, so the next agent finds the code as it is now.
Ask Korium
The questions every new session asks
| THE QUESTION YOU ASK | WHAT KORIUM ANSWERS FROM |
|---|---|
| "Why does the export skip canceled orders?" | The decision, who made it and when, linked to the test that holds it. |
| "Where is this function called, and what tests cover it?" | The code map: file, line span and test names, at the commit it indexed. |
| "What did we try last time this broke?" | Lessons and gotchas saved by earlier sessions in the same area. |
| "Does this change go against anything we decided?" | The decisions and principles on record for the code it touches. |
| "What is left from the last session?" | Its handoff note, in order. |
| "Who decided this, and when?" | Every decision carries who made it, the date and what it replaced. |
Example
Where, and why, in one answer
The agent gets the exact lines to read and the reason the code works that way. It reads the code itself before it changes anything.
Invoice export
Where
lib/billing/invoice_export.ex 41 to 97
test/billing/invoice_export_test.exs 88
Why
Finance disputed three canceled invoices in Q1, so canceled orders never reach the export. Lee decided, March 12.
Watch out
The April 2 change moved the filter from the view into the query. The old note is kept, marked replaced.
The map says where. The memory says why.
01 Fewer tokens
Read the right file first
The code map sends the agent to the lines that matter instead of every file that might.
02 Masthead is open source
The loop, packaged for Claude Code
Its hooks make every session search before it edits and save before it stops.
03 Any agent
Claude Code, Codex, any MCP client
One memory, whichever agent your team uses this month.