# I kept losing context every time a new Claude Code session started — worst when a reboot killed my terminal mid-task — retained record

> Retained reddit record from reddit:ClaudeCode.

Canonical: https://highsignal.app/data/records/8c5fd61b24cbe5df

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeCode)  published 2026-11-09 13:58 ingested 2026-11-10 02:42

[original source ↗](https://www.reddit.com/r/ClaudeCode/comments/1xrsc3n/i_kept_losing_context_every_time_a_new_claude/)

## retained text



```
I use Claude Code in the terminal, and every new session starts blank. That was fine until I started losing sessions I actually needed.

A reboot kills the terminal mid-task. Or I close the wrong window. Or tokens run out three hours into a fix and I come back to a fresh session. Either way the thread is gone — the stack, the bug, the four things I'd already ruled out.

The transcripts were on disk the whole time. Claude Code appends to `~/.claude/projects/<slug>/<session>.jsonl` as the session runs, so the one that died in the reboot is still sitting right there. I just had no way to ask it anything.

**"Just use --resume"**

I do, when I want the whole chat back. But `--resume` wants a session ID, and digging the right UUID out of a cluttered machine is its own chore. So now I ask for it first:

`recall the session id of my latest work on the checkout bug`

Then I resume with that. That one thing made `--resume` actually usable for me.

Most of the time though I don't want a whole conversation reloaded into context — I want one answer. And `--resume` only exists inside Claude Code.

**When tokens run out**

That's the other half of it. You hit the limit, move over to Codex or Cursor, and none of your Claude Code context comes with you. Steroid reads all four — Claude Code, Codex, Cursor, Copilot CLI — so the agent you switched to can answer from the session that just ran out.

# What it does

Three tools over the session files your agents already keep on disk:

`recall what I was doing on the checkout bug`

The answer cites agent, session, date and turn. `recall_turn` returns that full exchange and `recall_diff` returns the exact code change — both read the stored files directly, no model call. So you can check the answer against the source instead of trusting it.

You don't get the dead session back. You get the record of it, which turned out to be what I actually needed.

# Setup

One prompt pasted into Claude Code installs it and wires it up. No account. Free for individu

    Set up Steroid for me. It gives you, my coding agent, a memory of our past conversations.p of what you're about to do (a few short bullets, like a notification), then continuestraight into step 1 in the same reply:
    
    1. Check whether Steroid is installed: `steroid --version`, or `~/.local/bin/steroid --version`. If it is, go to step 3.
    2. Install it with the official installer from github.com/steroidkit/releases:
       - macOS / Linux: curl -fsSL https://raw.githubusercontent.com/steroidkit/releases/main/install.sh | sh
       - Windows (PowerShell): irm https://raw.githubusercontent.com/steroidkit/releases/main
       This shell may not see the new command yet, so use the path the installer prints. If the installer fails or refuses this machine, stop and show me why.
    3. Run Steroid's onboarding for this agent. If the `steroid-onboard` skill is already avanot (it was just installed, and you load skills at startup), run `steroidinstall-onboard-skill --json`, open the `SKILL.md` in the `steroid-onboard` folder it lists, and follow it.
    4. Finish with its status report and tell me to restart you so Steroid's tools load.

Rather do it yourself — install the binary, then run `/steroid-onboard` in Claude Code:

    curl -fsSL https://raw.githubusercontent.com/steroidkit/releases/main/install.sh | sh

No per-project setup. Recall in a new project and it indexes itself.

# What it actually puts on your machine

Given the week this sub has had with agent tooling, all of it:

* `~/.steroid/` — the app dir: `bin/`, `versions/`, `config.json`, `projects.json`, `logs/`,
* `~/.steroid/memory/<agent>/<project>/` — the index itself (`index.json`, `source.md`, `cache.json`, `diffs.json`, `meta.json`)
* A launcher on your PATH pointing at the binary in `~/.steroid/bin/` — on macOS a wrapper at
* One PATH line in your shell profile, only if that dir wasn't already there
* A per-user background daemon on your OS's user service manager — a LaunchAgent on macOS. Ustem daemon, no sudo
* The daemon runs two loops: every 5 minutes it checks indexed projects for new chat history and rebuilds incrementally only if the fingerprint changed; once a day it checks for a new release
* Indexing is lazy and per project — a project is indexed the first time you recall in it. Thdexes that already exist; it never scans your disk for repos
* One `steroid` MCP entry per agent you wire up — `~/.claude.json`, `~/.cursor/mcp.json`, `~/.codex/config.toml`, `~/.copilot/mcp-config.json`. Existing entries untouched
* Slash-command skills copied into each wired agent's skills dir (`~/.claude/skills/` etc.)
* A sentinel-wrapped block appended to the agent's global instructions file (`~/.claude/CLAUDE.md`, or the AGENTS.md equivalent) — removable without touching the rest
* All inside your home directory. `steroid uninstall` reverses it, including the PATH line, tries

What leaves the machine: the daily release check, and anonymous usage counts. `DO_NOT_TRACK=1sable` turns the counts off.

# What it can't do, before you ask

* It's a closed binary. `steroid daemon stop` stops the daemon, `steroid update --auto off` troid uninstall\` removes it. Updates are checksum-verified.
* The index is built and stored on your machine, no account, no server for recall — but the answer is written by whichever agent you ask, so that agent's provider sees the excerpts it uses.
* It indexes what Claude Code still keeps. The default 30-day cleanup still deletes transcrip those back — raise `cleanupPeriodDays` if you want more.
* Copilot CLI only. Copilot Chat in VS Code isn't covered yet.
* No published accuracy benchmark yet. That's exactly why every answer cites its source.

I'd like to know whether other CLI users hit the reboot thing, or whether I'm just the one ret.
```



document fetched 2026-11-10 00:39 · canonical https://reddit.com/r/ClaudeCode/comments/1xrsc3n/i_kept_losing_context_every_time_a_new_claude
