Claude Code memory, resume, and session retrieval are three different jobs

#claudecode #codex #sessionmanagement #developerworkflow

I used to think Claude Code's memory, /resume, and session history were all solving the same problem. They are not. Once I separated them, my workflow got a lot less fragile.

1. Memory is for reusable facts

CLAUDE.md, .claude/rules/, and auto memory are for information that should carry into future sessions. This is where project conventions, build commands, debugging notes, and repeated corrections belong.

Good use cases:

"Always use pnpm in this repo."
"The integration tests need Redis."
"API handlers live under this directory."
"Claude keeps missing this convention, so write it down."

This is not the same as reopening an old conversation. It is a way to give the next conversation better starting context.

2. Resume is for a known session

claude --continue reopens the most recent conversation in the current directory.

claude --resume reopens a specific session by ID or name, or lets you choose from a picker.

This is the right layer when you know which conversation you want to continue. Naming sessions helps a lot here.

Good use cases:

You stopped working on a PR yesterday and want to continue it.
You named the session auth-refactor and want to reopen it.
You only need the most recent conversation in this repo.

3. Retrieval is for when you forgot which session it was

The harder case is not "can Claude continue a known session?"

The harder case is "which old session was that?"

After a few repos, the ID, date, terminal tab, and exact session name blur together. You may remember the project, bug, or decision, but not the session handle. Digging through ~/.claude/projects/ raw JSONL files is not a workflow.

That is a separate retrieval problem:

browse old sessions by project
find the relevant conversation before reopening it
avoid guessing session IDs or scrolling through a flat picker
keep Claude Code and Codex work visible across repos

This is the layer I built Shelf for: https://github.com/Harukaon/shelf

Shelf is not a replacement for --continue, --resume, CLAUDE.md, or auto memory. Those native layers are useful. Shelf is meant to sit beside them as a desktop workspace for browsing and reopening Claude Code and Codex sessions by project.

Short version

Memory is for reusable facts.
Resume is for a known session.
Retrieval is for finding the old session when you do not remember the ID, name, tab, or date.

I keep the three layers separate now and it stopped most of the "where did that conversation go?" panic.