Cursor session format

Stores and files

Cursor IDE stores chats in state.vscdb under globalStorage/ and workspaceStorage/*/. The user root is ~/Library/Application Support/Cursor/User on macOS, %APPDATA%\Cursor\User on Windows and ~/.config/Cursor/User elsewhere; on Linux an existing $XDG_CONFIG_HOME/Cursor/User is used when available. DEJA_CURSOR_ROOT overrides IDE discovery.

Cursor CLI writes projects/<encoded-path>/agent-transcripts/**/*.jsonl below ${CURSOR_CONFIG_DIR:-~/.cursor}. DEJA_CURSOR_CLI_ROOT overrides transcript reads. Subagent transcripts are excluded unless DEJA_INCLUDE_SUBAGENTS=1 — for index

size, not because the parent already holds that work. Unlike Claude's sidechain

files, Cursor's have not been read here: what the parent keeps and what only the

child has is unverified for this harness, so deja indexes them as it finds them

rather than deriving an edge it has not seen.

IDE SQLite records

The cursorDiskKV(key, value) table has session values at composerData:<id> and message values at bubbleId:<composer-id>:<bubble-id>.

{"composerId":"composer-7","name":"Fix cache invalidation","createdAt":1784278800000,"lastUpdatedAt":1784278802000}
{"type":1,"text":"Why is this stale?","timestamp":1784278801000,"workspaceProjectDir":"/work/api"}

Bubble type: 1 maps to user; other numeric types map to assistant. Text uses text, falling back to rawText. Timestamps are Unix milliseconds. The first bubble with workspaceProjectDir supplies the project. Reading SQLite requires the sqlite3 CLI.

CLI JSONL records

{"role":"assistant","message":{"content":[{"type":"text","text":"The cache key omitted the locale."}]}}

Tool calls follow the Anthropic tool_use shape with Cursor's own names: path rather than file_path, Shell rather than Bash, old_string/new_string on StrReplace, and the whole file as contents on Write — the last of these is the only record a created file's lines were ever in a session.

Only user and assistant roles are retained. Content follows the Anthropic string-or-parts shape. Control records such as turn_ended are ignored. Records carry no timestamp field; a user turn's text starts with a <timestamp>Sunday, Jul 26, 2026, 1:06 PM (UTC+3)</timestamp> block, which deja reads and carries forward to the assistant turns that answer it. A transcript with no readable block falls back to the file's modification time.

The second CLI store

Cursor CLI also writes one SQLite store per chat:

~/.cursor/chats/<workspace-hash>/<chat-uuid>/meta.json
~/.cursor/chats/<workspace-hash>/<chat-uuid>/store.db

store.db holds blobs(id TEXT, data BLOB) and meta(key, value). The blobs are content-addressed and the tree needs no schema to walk: meta's single value is hex-encoded JSON naming latestRootBlobId, that blob is protobuf whose repeated field 1 is a list of 32-byte child digests in message order, and each child is plain JSON in the Vercel AI SDK shape — {"role","content"} with text, reasoning, tool-call and tool-result parts, the calls named as above but keyed toolName/args.

Each chat also has a meta.json beside its store.db holding the cwd it ran in, and the <workspace-hash> above it is md5 of that path; deja reads it for the project name and the resume directory (#4193). The transcript holds every turn the store does except the <user_info> environment preamble, but none of the tool results, so deja reads the store for those alone: each transcript call is paired with the store's call of the same tool name and arguments, in order, and takes its result. A Shell result gives the exit status on the command and the printed output without Cursor's Exit code / Command output framing; the results of GetMcpTools and GetDynamicTools, which are the MCP servers' own descriptions, are skipped. Every chat with a store.db had a transcript beside it, and deja doctor reports the count of chats with none — silent today, and the only signal if a release stops writing agent-transcripts.

Wiring

deja install cursor adds the server to ${CURSOR_CONFIG_DIR:-~/.cursor}/mcp.json, writes the shared skill ~/.agents/skills/deja-history/SKILL.md and the /deja command in commands/deja.md beside it. deja install cursor-auto adds the same plus hooks.json entries: sessionStart (deja hook-context), beforeSubmitPrompt (hook-prompt, interactive TUI only — headless -p skips it), preToolUse (matcher ^(Shell|Write|Task)$), postToolUse and postToolUseFailure (matcher ^Shell$; the second fires for a tool that errors, with the text in error_message), preCompact and sessionEnd (hook-session-end, so the session you closed is back in the next one's MCP recall). Cursor also runs the hooks in ~/.claude/settings.json and dedupes them against its own by exact command string, so a machine with both wired gets one injection. Cursor tests a matcher as a regex against tool_name for the three tool events (2026.09.02), so deja does not start on every read and grep.

Resume

CLI chats only. A transcript is named after the chat id cursor-agent --resume

takes, and the command runs in the project directory because Cursor lists chats

per workspace — it looks a chat up under chats/<md5 of the cwd>/<id> — so deja

prints cd <project> && cursor-agent --resume <id>, the directory read from the

chat's meta.json, whose folder is that md5. The encoded transcript folder is

only the fallback: it blanks a dot, a space or a non-ASCII character and cuts a

long path with a hash, so it cannot be read back for those. A chat whose

directory is gone, is not recorded, or is no longer in cursor-agent's store is

refused with deja show instead — cursor-agent --resume with no id lists

chats from every directory. Live-verified: the resumed chat answered from its own

history. IDE chats carry a composer id from state.vscdb that the CLI does not

take, so those still reopen only in the editor.

Known quirks and drift

  • Cursor moved modern IDE chats toward global storage while older versions used workspace databases; both are scanned.
  • SQLite values are JSON inside a key-value table and malformed or null entries occur.
  • CLI project path encoding has the same separator-versus-hyphen ambiguity as Claude Code.
  • A parsed message is capped at 1 MiB. CLI subagent transcripts are opt-in.
  • Both CLI layouts are written by cursor-agent 2026.09.02-c22c1a3, minutes apart in the same session.

Last verified: 2026-10-07

deja reads this format and thirty-seven others, and turns what it finds into memory your agents can search. See the harness matrix for what is wired where, or install it and search your own history.

Found this useful? Star deja-vu on GitHub.