Git Reflog: Recover Lost Commits Without Moving Your Branch
In the reset lab below, HEAD@{1} points to the commit you already have. The wanted commit is three entries back. Two branch switches are enough to make a copied recovery command select the wrong snapshot.
To recover lost commits with Git reflog, inspect the local record, check the candidate's contents, and create a new branch at its full commit ID. Plain branch creation leaves your current branch, staged changes, and working files in place. Once the commit has a name, you can decide how to use it.
The three cases cover a commit left behind by reset, the original commit replaced by amend, and work made in detached HEAD. Each starts in its own disposable repository and ends with a rescue branch. The reported results come from executed checks, with the environment and limits recorded below.
Find, inspect, then give the commit a name
Start in the local repository where you made the commit. Read its current status and the records for HEAD and the branch involved:
git status --short
git reflog show HEAD
git reflog show main
Replace main with your branch name. If that branch or its reflog doesn't exist, the last command can fail. Continue inspecting HEAD's available record; don't use a selector for a branch Git couldn't find.
A branch name points to a commit. Its reflog records earlier values of that reference when logging is enabled. HEAD's record also includes branch switches. These local reference histories differ from the commit ancestry shown by an ordinary git log. See the Git reflog manual.
| Expression | What it selects |
|---|---|
HEAD~1 |
The first parent of the current commit |
'HEAD@{1}' |
The previous recorded value in HEAD's reflog |
'main@{1}' |
The previous recorded value in main's reflog |
The number in @{1} counts reflog entries, not parents. A branch switch can add an entry even when both branches point to the same commit. Later recorded updates shift the numbers, so quote selectors in the shell and resolve the selected entry before doing more work. Git's revision syntax reference explains these forms and their quoting requirements.
For example, only if the record you inspected identifies main@{1} as the candidate, capture its full ID and examine it:
candidate=$(git rev-parse 'main@{1}')
git --no-pager show --no-patch --format=fuller "$candidate"
git --no-pager show --stat "$candidate"
git show "${candidate}:notes.txt"
Replace notes.txt with a relevant path listed in the commit. These commands show the commit metadata, a change summary, and the committed version of that file. They don't replace today's file. Use the message to locate plausible candidates, then check the contents that distinguish the wanted snapshot. For a larger recovery, inspect the other relevant files too. The git show manual covers these inspection forms.
Once you've confirmed the commit, retain it with a new, unused branch name:
git branch rescue-found-commit "$candidate"
Plain git branch creates the reference without checking it out, as the branch manual specifies. If the name already exists, inspect it and choose a different name; don't add -f out of habit. The captured ID keeps this operation tied to the object you inspected, even if reflog numbering changes later.
At this point the commit is reachable through rescue-found-commit. Its changes haven't been applied to your current files. That completes the preservation step.
Set up a disposable recovery lab
The three scenarios were executed on October 24, 2026, using Git 2.54.0 (Apple Git-157), macOS 26.6.1, arm64. The checks compared the current branch and commit, the index and working versions of notes.txt, repository status, and the reset case's untracked file before and after branch creation. These results concern the sample repositories; they don't establish the state of a repository you're recovering.
Define these two functions once in a Bash or zsh session. Each case calls the setup inside parentheses to create a fresh temporary repository. The parentheses also keep directory changes and shell options inside that subshell. set -eu stops the case on a command error or unset variable. If a command fails, inspect the error and start the case again in a fresh lab.
new_reflog_lab() {
lab_dir=$(mktemp -d /tmp/git-reflog-practice.XXXXXX)
cd "$lab_dir"
printf 'Lab directory: %s\n' "$lab_dir"
git init --initial-branch=main
git config user.name "Git Practice"
git config user.email "practice@example.com"
git config commit.gpgsign false
git config core.hooksPath /dev/null
git config core.logAllRefUpdates true
printf 'one\n' > notes.txt
git add notes.txt
git commit -m A
printf 'two\n' > notes.txt
git add notes.txt
git commit -m B
printf 'three\n' > notes.txt
git add notes.txt
git commit -m C
}
current_lab_state() {
git symbolic-ref --short HEAD || return
git rev-parse HEAD || return
git log -1 --format=%s || return
git show :notes.txt || return
cat notes.txt || return
git status --porcelain=v1
}
The settings apply only to the temporary repository. Every setup creates A → B → C: notes.txt contains one in A, two in B, and three in C. The branch main starts at C with a clean index and working tree. Your commit IDs will differ from the recorded run.
The state helper prints, in order, the branch, full HEAD ID, commit subject, index contents, working contents, and short status. Each case saves this output before and after rescue and checks that the two values match. A successful test command prints nothing; a failed comparison stops the case.
Run the reset and amend commands below only in these disposable labs. They deliberately move or replace sample history. For the differences between undo operations themselves, use the separate Git reset, restore, and revert practice.
Case 1: Recover a commit after reset, with new edits already present
This case moves main back to A, switches to another branch at A, then returns to main. It then stages staged, edits the working file to working, and adds an untracked file containing keep me. Before running it, predict what HEAD@{1} will select and where you'll find C.
(
set -eu
new_reflog_lab
git branch side HEAD~2
git reset --hard HEAD~2
git switch side
git switch main
printf 'staged\n' > notes.txt
git add notes.txt
printf 'working\n' > notes.txt
printf 'keep me\n' > scratch.txt
git reflog show --format='%h %gd %gs' HEAD
git reflog show --format='%h %gd %gs' main
git show -s --format=%s 'HEAD@{1}'
candidate=$(git rev-parse 'main@{1}')
git --no-pager show --no-patch --format=fuller "$candidate"
git show "${candidate}:notes.txt"
before=$(current_lab_state)
scratch_before=$(cat scratch.txt)
git branch rescue-reset "$candidate"
after=$(current_lab_state)
scratch_after=$(cat scratch.txt)
test "$before" = "$after"
test "$scratch_before" = "$scratch_after"
printf '%s\n' "$after"
cat scratch.txt
git show rescue-reset:notes.txt
)
The observed entries were:
| Reflog entry | Commit | Recorded event |
|---|---|---|
HEAD@{0} |
A | Switch from side to main |
HEAD@{1} |
A | Switch from main to side |
HEAD@{2} |
A | Reset to HEAD~2 |
HEAD@{3} |
C | Commit C |
main@{0} |
A | Reset to HEAD~2 |
main@{1} |
C | Commit C |
HEAD@{1} produced subject A. C was at HEAD@{3} and main@{1} in this sequence. The switches added entries to HEAD's record while main remained at A, so main's reflog stayed shorter.
Candidate inspection printed C's subject and file contents three. After creating rescue-reset, both comparisons passed:
| Location | Before rescue | After rescue |
|---|---|---|
| Current branch / commit | main / A | main / A |
Index notes.txt |
staged |
staged |
Working notes.txt |
working |
working |
| Status | MM notes.txt, ?? scratch.txt |
Same |
Untracked scratch.txt |
keep me |
keep me |
New branch rescue-reset |
Absent | C, containing three |
The branch rescue-reset now retains C and its ancestors, including B. Main still points to A, and today's sample edits are intact. Reading the rescued snapshot doesn't require checking it out.
Case 2: Recover the original amended commit
Amending creates a replacement tip commit, as described in the commit manual. To recover the original, look for the branch's position before the amendment and verify the old contents.
Start a fresh lab at C, change notes.txt to revised, then amend:
(
set -eu
new_reflog_lab
printf 'revised\n' > notes.txt
git add notes.txt
git commit --amend -m 'C revised'
git reflog show --format='%h %gd %gs' main
candidate=$(git rev-parse 'main@{1}')
git --no-pager show --no-patch --format=fuller "$candidate"
git show "${candidate}:notes.txt"
before=$(current_lab_state)
git branch rescue-before-amend "$candidate"
after=$(current_lab_state)
test "$before" = "$after"
printf '%s\n' "$after"
git show HEAD:notes.txt
git show rescue-before-amend:notes.txt
)
Here main@{0} was C revised, and main@{1} was the original C. Inspection printed three for the candidate, distinguishing it from the amended snapshot's revised.
The comparison passed. main stayed at C revised, the index and working file stayed revised, and status remained clean. Reading HEAD:notes.txt printed revised; reading rescue-before-amend:notes.txt printed three.
Both snapshots now have branch names. Creating rescue-before-amend hasn't undone the amendment. You can compare the versions before deciding what to use; any later change to shared history needs coordination with the people using it.
Case 3: Find a detached commit in HEAD's reflog
A detached HEAD points directly to a commit. Committing in that state moves HEAD without moving an existing branch to the new commit. The git switch manual explains detached operation and how a branch can retain that work.
This time, detach at B, create D with contents detached, then return to main at C:
(
set -eu
new_reflog_lab
git switch --detach HEAD~1
printf 'detached\n' > notes.txt
git add notes.txt
git commit -m 'D detached'
git switch main
git reflog show --format='%h %gd %gs' HEAD
git reflog show --format='%h %gd %gs' main
candidate=$(git rev-parse 'HEAD@{1}')
git --no-pager show --no-patch --format=fuller "$candidate"
git show "${candidate}:notes.txt"
before=$(current_lab_state)
git branch rescue-detached "$candidate"
after=$(current_lab_state)
test "$before" = "$after"
printf '%s\n' "$after"
git show HEAD:notes.txt
git show rescue-detached:notes.txt
)
In this sequence, HEAD@{0} recorded the return to main, and HEAD@{1} identified D. Main's reflog contained C, B, and A. It had no D entry because main had never pointed to D.
The candidate's subject was D detached, and its file contained detached. After rescue, the comparison passed: the current branch remained main at C, its index and working file remained three, and the repository stayed clean. rescue-detached pointed to D.
Here HEAD@{1} finds the wanted commit. In case 1 it finds A. The useful habit is inspecting the record and contents; the right number depends on what happened in that repository.
What reflog recovery can and can't retain
This workflow needs both a useful reflog entry and its commit object. It recovers committed snapshots still available in the same local repository. It doesn't recover never-committed edits merely because a hard reset or restore discarded them: reflogs record reference positions, not every edit to the working tree.
Reflogs are local. Pushing or cloning a repository doesn't transfer these recovery records to another clone. Another machine may have useful commits of its own, but it doesn't automatically have your local reflog.
Git documents default reflog expiration settings of 90 days generally and 30 days for entries unreachable from the current tip. The settings are gc.reflogExpire and gc.reflogExpireUnreachable; configuration and explicit commands can change them. They are not guaranteed recovery windows. Expiration removes reflog entries, while pruning removes eligible unreferenced objects. A commit object may remain after its entry expires, even though you can no longer find it through that entry. See the garbage-collection manual for these defaults and object-retention rules.
A rescue branch gives the commit a local reference, retaining it through ordinary reachability rules while that branch remains. It isn't an off-machine backup. Losing the repository's Git data can still lose the commit and the rescue branch together.
After preservation, choose the integration step separately. Switching to the rescue branch changes the checkout; resetting the current branch to the recovered commit moves that branch; applying a commit's changes is another operation. Each has different effects on current work. The recipe above deliberately ends once the wanted commit has a branch name.
Check your reasoning without the terminal
Close the terminal and answer these from the three cases:
- After the reset and two switches, what does
HEAD@{1}select? Where is C? - Does creating
rescue-resetputthreeinto the working file? - How can you read the original C after amend without undoing the amendment?
- Why is D absent from main's reflog in the detached case?
- Why capture a full commit ID instead of continuing to use the reflog selector?
Then check your reasoning:
| Question | Answer |
|---|---|
| 1 | A: the recorded switch to side also pointed to A. C is at HEAD@{3} and main@{1}. |
| 2 | No. Branch creation adds a reference at C without checking it out. The file stays working, and the index stays staged. |
| 3 | Run git show rescue-before-amend:notes.txt. It prints three; main's amended snapshot still contains revised. |
| 4 | D was created while HEAD was detached. Main remained at C and never recorded D as its value. HEAD's reflog recorded the detached work. |
| 5 | Later reference updates can shift the entry numbers. The full ID keeps identifying the object you inspected. |
After a gap, redraw the branches and answer again without looking. If a selector surprises you, rerun that case in a fresh lab and explain which reference moved at each step.
The English Git Commands Flashcards deck has 98 cards, including reflog inspection and creating a recovery branch. Use those cards alongside the state predictions above: recall the command, then explain what it changes. Regular recall and spaced review give you repeated chances to retrieve that knowledge after a gap, helping long-term retention. They don't guarantee permanent memory or replace inspecting a real repository before changing it.
Production note: AI tools performed substantial research, drafting, and editing for this article. An AI agent executed the disposable Git cases behind the reported results. The accompanying illustration is AI-generated. These sample checks aren't a claim of personal testing or human review by Kirill.