Recallium works best when memories preserve complete units of reasoning and stay connected to the work they explain.
1. Search before editing established code
Ask what the team already knows before changing an important file or behavior:
What context should I know before editing src/auth/tokens.ts?
Earlier decisions, warnings, and bug fixes can change the safe implementation path.
2. Capture complete units, not narration
Save one decision, root cause, experiment, or operational procedure at a time. Include enough context for a future agent to use it without the original chat.
Avoid saving a stream of tiny updates such as “opened the file” or “changed one line.”
3. Preserve the reason
“We chose Postgres” is incomplete. Record why it won, which alternatives were considered, and what trade-off the team accepted.
The code shows what exists. Recallium should preserve why it exists.
4. Connect the files
Ask your agent to link the files involved in a decision or fix:
Save this token rotation fix and include every file we changed or inspected.
File links make the context discoverable when a later session returns to the same code.
5. Save hard-won context while it is fresh
Capture a meaningful decision, diagnosis, or result when it lands. If work stops unfinished, leave a clear continuation point with what remains and the next action.
A useful test is whether a teammate could understand the result two weeks later using only the memory and the repository history.
