Where AI coding tools store your chat history.

Cursor is a VS Code fork, so it inherits VS Code's storage layout: one large SQLite database under globalStorage, plus a workspaceStorage/<hash>/ directory per project folder. The per-workspace databases carry a workspace.json naming the folder and a composer.composerData key listing that workspace's conversations. The conversation content itself is in the global database.

Almost all of it sits in one table, cursorDiskKV, as JSON blobs keyed by id. The database is large and grows: on the machine this was measured, roughly 2.5 GB with a further ~1.5 GB write-ahead log alongside it. Anything that copies the file before reading copies all of that, every time.

There is a second, tempting file: conversation-search.db. It is smaller, tidier, and already full-text indexed. It is also a derived cache — on the same machine it held 330 conversations against 403 in the real store, and each one is flattened into a single blob with the speaker labels discarded, so you cannot tell what you said from what the model said. Read the real store.

One SQLite detail costs an afternoon: cursorDiskKV.value is declared BLOB, but Cursor writes JSON into it as TEXT. SQLite applies no type affinity to a BLOB column, so a reader that asks for bytes gets a type error on every single row. Combined with an iterator that silently drops errors, that produces an empty result set and no failure — COUNT(*) reports thousands of rows while SELECT value appears to return nothing.

The most readable of the six. Each session is a JSON Lines file — one JSON object per line, appended as the conversation happens — inside a directory named after the project it ran in. A path like /Users/you/code/api becomes the directory -Users-you-code-api.

Because it is append-only plain text, it is the one store you can usefully grep directly, and the one that survives being copied while the tool is running.

It is also the only one of the six that deletes itself. Claude Code removes session files older than cleanupPeriodDays — 30 by default — at startup, with no prompt.

Sessions are JSONL like Claude Code's, but partitioned into year/month/day directories, with a parallel ~/.codex/archived_sessions tree for older ones. Both need reading if you want everything.

The trap is identity. A resumed run or an approval prompt spawns a child thread that carries the same session_id as its parent while living in its own file with its own id. Anything that treats session_id as the primary key will map two files onto one record, and each pass will overwrite the other — an indexer doing this reindexes the same pair on every run, forever, with no error to notice.

Zed keeps agent threads in one SQLite database under its application support directory. Unlike the VS Code forks there is no workspace-keyed split; the database is the whole store.

Zed's Windows support is recent and its store location there was never confirmed on a real machine, so anything reading this cross-platform should probe both the roaming and local app-data directories rather than assume one.

Kiro inherits the VS Code layout, which makes it easy to assume its conversations are inside state.vscdb with everything else. They are not. The agent writes ordinary files under globalStorage/kiro.kiroagent/workspace-sessions/, with a sessions.json acting as the index.

That makes it the most straightforward of the VS Code forks to read: no SQLite, no blob columns, no cache to mistake for the real thing.

Antigravity splits one conversation across three files, and the one that looks canonical is the one you cannot read. conversations/<uuid>.pb is encrypted: measured at 7.9997 bits per byte of entropy, which is indistinguishable from random data, with no recoverable strings.

The same conversation is also written, unencrypted, to an operational log at brain/<uuid>/.system_generated/logs/overview.txt — JSONL, one step per line, with user messages, model replies and timestamps. Titles, workspace folders and branches live in a third file, agyhub_summaries_proto.pb, a binary protobuf with no published schema.

Two practical notes. Not every conversation has a log — on the machine this was measured, 11 of 13 did, and the two without were the oldest. And Antigravity truncates its own tool arguments at 256 bytes, appending a note after a raw newline; since those arguments are double-encoded JSON strings, the result is a string that is both unterminated and carrying a control character illegal inside JSON. A parser that does not repair that before parsing fills up with stray escape sequences and looks like it worked.