You worked this out already.
braids finds where.

Find itOne field over everything you have

Every message and tool call, every memory a project keeps, and the name of every file a session left on disk. Each hit says which kind it is, and ↵ opens the screen that can do something about it: a conversation at that turn, a memory in the reader, a file in the browser.

$ braidsthen / and a word

Scopes, kinds, dates, and how the ranking works

ExplainWhich conversation produced this file

git knows when a file changed. braids knows what was being said in that directory at the time. Joining them gets you from a line of code back to the conversation it came out of.

~/src/storefront/checkout/handler.go
2 commits, against 6 conversations
  6 ran in this repository

996282d5  2026-09-04 13:55  checkout: defer the tax lookup out of the lock
    Checkout stalls at 80% under load, then recovers on its own  ·  storefront  ·  24 turns in the window
      last said 4 minutes before, at turn 24:
        so we need the deferral to rejoin before receipt
      braids --lane a1b2c3d4
    Option C: move the lock inside the tax client and re-run the load test  ·  storefront  ·  6 turns in the window
      last said under a minute before, at turn 6:
        stall is gone and the receipt total is right
      braids --lane b2c3d4e5

5b407b53  2026-09-02 14:05  checkout: take the tax quote inline
    Blobstore is doing 11 KiB reads per object instead of batching them  ·  storefront  ·  4 turns in the window
      last said 7 minutes before, at turn 4:
        the reader asks per object, not per batch
      braids --lane d4e5f6a7

These conversations were live when those commits landed. braids does not
know that they caused them; it knows when things happened and where.

It does not claim those conversations caused those commits, and it says so outright. It offers you where to look. That is all a git log and two columns can honestly give you, and it is why this costs neither a model nor a moment.

Every command braids has

Then carry onStart a new conversation from any turn

Finding the moment is half of it. The spine collapses a long conversation to its landmarks: what you asked, what came back, where a tool call failed, where the context got compacted, where each branch left.

$ braidsthen ↵ on a conversation

Put the cursor on a turn and press b. braids writes a new session file holding that turn's ancestry, then hands you the claude --resume command. Ask for a workspace and the branch gets its own git worktree, so two branches can edit the same file and never collide. The transcript you branched from is opened read only, and a test hashes it before and after to prove it.

Branching, merging and promoting a subagent

Or let ClaudeIt can search its own past

Every command that reports something takes --json, with whole IDs meant to be passed back. Install the skill and Claude Code reaches for your history itself, at the moments worth reaching: when you refer to earlier work, when you ask why something is the way it is, and before it proposes something that may already have been tried.

braids skill --install
# or, from inside Claude Code
/plugin marketplace add Ashes47/braids

Most of what the skill teaches is when not to look. Searching your history on every question is slow and usually beside the point, so it names the cases the code in front of it already answers. And it never claims a past conversation caused anything: it quotes what was said, with the turn number, and leaves the conclusion to you.

Everything an agent can read

The mapAnd the shape of all of it

One row per conversation, with branches under the conversation they were cut from, tagged with the turn they left at. Search is the front door; this is where you see what you have and who is waiting on you.

$ braids

• your turn means the session asked you something and stopped. n jumps to the next one, which is a decent way to start a morning.
• The indent is lineage. t14 on a branch is the turn of the parent it was cut at.
• Every screenshot here is real output, colours and all, captured against a fake ~/.claude.

Every key on the map and the spine

And the restWhat else is in there

Five more screens, each with a page of its own. They are the parts of a session you cannot see from inside it.

LocalIt reads your transcripts, so this matters

• No network calls. There is no HTTP client and no listener in it. Run go list -deps ./cmd/braids | grep net/http and see for yourself.
• Nothing leaves the machine. The index is one SQLite file at ~/.braids/index.db.
• What it writes stays yours. ~/.braids is 0700 and the index is 0600, because it holds the full text of every message it has read.
• It never writes to a transcript the harness owns. Branching, merging and promoting all write new files.
• Delete braids and you lose a view, not a conversation. All of it is derived from ~/.claude.

What it reads, writes, and never writes

NumbersMeasured on one laptop, against a real corpus

28 conversations, about 62,000 messages, about 360 MB of transcripts. Rounded, because a corpus you are still working in will not sit still.

That last row is the honest cost: about a third of the transcripts, because searching text means keeping a copy of it. Two thirds of the index is that copy and the search structure over it.

InstallOne binary, and one way out

# picks the build for your machine and checks it against the published sums
curl -fsSL https://braids.chat/install.sh | sh

# or with Go
go install github.com/Ashes47/braids/cmd/braids@latest

# then
braids index   # read every transcript under ~/.claude
braids         # open the map

Run that first line again to update. It replaces braids where it already lives rather than leaving a second copy on your PATH, and says already installed and stops when there is nothing to do.

Staying current

braids cannot tell you a newer version exists, because finding out would mean asking, and it makes no network calls. What it can say is how long it has been since anyone checked. After a month the map's header adds 3 months old · v to update to the version row, and v runs the installer above, printing the command first. Checking and finding yourself current resets the clock, so it asks once a month at most and never claims you are behind.

Uninstalling

# take the hook back out of ~/.claude/settings.json, if you added it
braids hooks --remove

# then the binary and the index
rm "$(command -v braids)"
rm -rf ~/.braids

That is all of it. Your conversations live in ~/.claude and are never touched, so removing braids costs you the map and nothing else.

Install, first index, and a five minute tour

If it is usefulStar the repo

braids is one person's tool, made public because the problem is not one person's. A star is how the next person finds it. An issue is how it gets better.