ChatGPT desktop

agentgrep searches ChatGPT desktop’s conversation titles and scheduled-task prompts on Windows. Neither app keeps message text on disk in a form another program can read: the Windows app caches only the conversation list and the scheduled-task list, and the macOS app encrypts its conversation files with a key that only OpenAI’s own code can read.

Base paths: %LOCALAPPDATA%/Packages/OpenAI.ChatGPT-Desktop_2p2nqsd0c76g0/LocalCache/Roaming/ChatGPT (Windows, reached from WSL through each Windows profile) and ~/Library/Application Support/com.openai.chat (macOS).

observed_version: ChatGPT desktop (observed 2026-11-14).

What is searched

chatgpt.conversation_list reads the Windows app’s Chromium Local Storage, a LevelDB directory, without a LevelDB library and without taking the app’s lock. The app caches the first page of its conversation list there (about 28 conversations, plus project lists), so agentgrep finds each cached conversation by title. Each title becomes one record whose session id is the ChatGPT conversation id, with the creation time, project id, and archive, star, and temporary-chat flags in its metadata. Only the conversation-list keys are decoded; the store’s session data and secrets are never read.

chatgpt.http_cache reads the app’s Chromium HTTP cache, in Chromium’s blockfile format, the same way: without Chromium and without its lock. Three cached responses are read: the conversation list, each project’s conversation list (project conversations the Local Storage list leaves out), and the scheduled-task list. Conversations become title records as above, and each scheduled task with a prompt becomes a user prompt record holding the text the user scheduled; a task’s model output is not read. Every other cached response, account and session data among them, is never read past its URL. The bodies are Brotli-compressed, which the brotli package decodes.

Both stores count as conversation content, so a conversation-scope or broad-scope search reads them (a conversation scope runs at exhaustive effort unless told otherwise; targeted effort does not read the list):

$ agentgrep search --agent chatgpt --scope conversations deploy

A result looks like this in --json output (values are illustrative):

{
  "kind": "history",
  "agent": "chatgpt",
  "store": "chatgpt.conversation_list",
  "path": "%LOCALAPPDATA%/Packages/OpenAI.ChatGPT-Desktop_2p2nqsd0c76g0/LocalCache/Roaming/ChatGPT/Local Storage/leveldb/CURRENT",
  "text": "Plan the widget launch",
  "title": "Plan the widget launch",
  "role": "assistant",
  "timestamp": "2026-09-25T10:15:02.123456Z",
  "session_id": "68d1c0aa-0000-0000-0000-000000000000",
  "metadata": {"created_at": "2026-09-12T08:00:00.000000Z", "starred": true}
}

The role is assistant because ChatGPT writes a conversation’s title itself; the cache does not mark a title the user renamed, so role:user leaves the titles out.

The printed path works as a path: query, so the ChatGPT titles can be picked out of a broader search:

$ agentgrep search --scope all 'path:%LOCALAPPDATA%/Packages/OpenAI.ChatGPT* deploy'

The cache holds the conversations the sidebar showed when the app last ran; older conversations and every message body stay on OpenAI’s servers. Because it is a cache, it can still list a conversation deleted or moved off the first page since the app last refreshed: agentgrep keeps every conversation any cached snapshot lists, newest title per conversation.

Stores

Role:

primary_chat

Format:

opaque

Coverage:

catalog_only

Adapter:

chatgpt.file_metadata_summary.v1

catalog store
catalog store
Role:

supplementary_chat

Format:

opaque

Coverage:

catalog_only

Adapter:

chatgpt.file_metadata_summary.v1

Role:

app_state

Format:

leveldb

Coverage:

default_search

Adapter:

chatgpt.local_storage_conversation_list.v1

default store
default store
Role:

app_state

Format:

chromium_cache

Coverage:

default_search

Adapter:

chatgpt.http_cache_json.v1

catalog store
catalog store
Role:

cache

Format:

opaque

Coverage:

catalog_only

Adapter:

chatgpt.file_metadata_summary.v1

chatgpt.conversations

chatgpt.conversations
catalog store
catalog store
chatgpt.conversations

macOS app, one .data file per conversation (898 B to 349 KB). The app (bundle com.openai.chat) seals each file with CryptoKit AES-GCM under a 256-bit key and stores a property list inside; the key is a data protection keychain item in OpenAI’s team access group, which macOS opens only to code signed with that entitlement, so no other program can decrypt the files. Listed for inventory as metadata only; never searched.

Agent:

ChatGPT

Role:

primary_chat

Format:

opaque

Coverage:

catalog_only

Path:

${HOME}/Library/Application Support/com.openai.chat/conversations-v3-<uuid>/<UUID>.data

Adapter:

chatgpt.file_metadata_summary.v1

Data version:

chatgpt.opaque.v1

Version strategies:

-

Observed:

ChatGPT desktop (2026-11-14)

Default search:

no

chatgpt.drafts

chatgpt.drafts
catalog store
catalog store
chatgpt.drafts

macOS app, unsent prompt drafts (183 B to 1.8 KB), encrypted like the conversation files. Listed for inventory as metadata only; never searched.

Agent:

ChatGPT

Role:

supplementary_chat

Format:

opaque

Coverage:

catalog_only

Path:

${HOME}/Library/Application Support/com.openai.chat/drafts-v2-<uuid>/<UUID>.data

Adapter:

chatgpt.file_metadata_summary.v1

Data version:

chatgpt.opaque.v1

Version strategies:

-

Observed:

ChatGPT desktop (2026-11-14)

Default search:

no

chatgpt.conversation_list

chatgpt.conversation_list
default store
default store
chatgpt.conversation_list

Windows app, Chromium Local Storage in LevelDB (*.log write-ahead logs and Snappy-compressed *.ldb tables). Keys cache/user-<account>/<id>/conversation-history and .../conversation-history-without-projects hold the cached first page of the conversation list as JSON: {timestamp, value: {pages: [{items, total, limit, offset}]}, version}, each item with id, title, create_time and update_time (ISO 8601), gizmo_id (project), and archive, star, and temporary flags; mapping and snippet are always null, so the list carries titles, not messages. The same store holds session data and secrets under other keys, which are never decoded.

Agent:

ChatGPT

Role:

app_state

Format:

leveldb

Coverage:

default_search

Path:

%LOCALAPPDATA%/Packages/OpenAI.ChatGPT-Desktop_2p2nqsd0c76g0/LocalCache/Roaming/ChatGPT/Local Storage/leveldb/

Adapter:

chatgpt.local_storage_conversation_list.v1

Data version:

chatgpt.local_storage_conversation_list.v1

Version strategies:

-

Observed:

ChatGPT desktop (2026-11-14)

Default search:

no

One title record per conversation, newest snapshot per id; its session id is the ChatGPT conversation id. Exhaustive or broad-scope search reads it; prompt search does not.

chatgpt.http_cache

chatgpt.http_cache
default store
default store
chatgpt.http_cache

Windows app, Chromium blockfile HTTP cache (index, data_0 to data_3, external f_<n> files); bodies are content-encoding: br. Three cached API responses are read: /backend-api/conversations ({items, total, limit, offset}, items as in chatgpt.conversation_list), /backend-api/gizmos/<project>/conversations ({items, cursor}, project conversations the Local Storage list leaves out), and /backend-api/tasks ({tasks, cursor}; each task’s prompt is the text the user scheduled, messages and final_message are model output and are not read). Other entries, including /api/auth/session and account responses, are never read past their URL.

Agent:

ChatGPT

Role:

app_state

Format:

chromium_cache

Coverage:

default_search

Path:

%LOCALAPPDATA%/Packages/OpenAI.ChatGPT-Desktop_2p2nqsd0c76g0/LocalCache/Roaming/ChatGPT/Cache/Cache_Data/

Adapter:

chatgpt.http_cache_json.v1

Data version:

chatgpt.http_cache_json.v1

Version strategies:

-

Observed:

ChatGPT desktop (2026-11-14)

Default search:

no

Title records keyed by conversation id, newest snapshot per id, and a user prompt record per scheduled task with a prompt. Exhaustive or broad-scope search reads it; prompt search does not.

chatgpt.web_profile

chatgpt.web_profile
catalog store
catalog store
chatgpt.web_profile

Windows app, an Electron shell around chatgpt.com installed as an MSIX package. Windows redirects its %APPDATA% writes to the package’s LocalCache/Roaming, so from WSL the data is only under %LOCALAPPDATA%/Packages. The directory is a Chromium profile (IndexedDB, Local Storage, Session Storage, Cache, Preferences, a two-byte config.json). Decoded, the IndexedDB ConversationsDatabase.conversations store holds no records and no conversation body response is cached, so conversation text is held server-side; the conversation list is read from Local Storage (chatgpt.conversation_list). Listed for inventory as metadata only; never searched.

Agent:

ChatGPT

Role:

cache

Format:

opaque

Coverage:

catalog_only

Path:

%LOCALAPPDATA%/Packages/OpenAI.ChatGPT-Desktop_2p2nqsd0c76g0/LocalCache/Roaming/ChatGPT/IndexedDB/https_chatgpt.com_0.indexeddb.leveldb/

Adapter:

chatgpt.file_metadata_summary.v1

Data version:

chatgpt.opaque.v1

Version strategies:

-

Observed:

ChatGPT desktop (2026-11-14)

Default search:

no

What is not searched

chatgpt.conversations and chatgpt.drafts are the macOS app’s conversations-v3-<uuid>/<UUID>.data and drafts-v2-<uuid>/<UUID>.data files. The app (bundle com.openai.chat) seals each file with CryptoKit AES-GCM under a 256-bit key and stores a property list inside. The key is a keychain item in OpenAI’s team access group in the data protection keychain, which macOS opens only to code signed with that entitlement (TN3137: On Mac keychain APIs and implementations), so neither agentgrep nor the security tool can decrypt the files. These findings come from the app’s binary and Apple’s documentation, not from a Mac.

chatgpt.web_profile is the rest of the Windows profile. Decoded, its IndexedDB ConversationsDatabase holds no records. Its HTTP cache holds conversation-list responses, which repeat the Local Storage list, and scheduled-task records, whose prompts are the one piece of typed text in the profile, but no conversation body. The cache entries are Brotli-compressed and evicted as the app runs, and agentgrep ships no Brotli decoder, so the cache is listed, not read.

find and the agentgrep://sources resource list these stores so a storage audit shows that the files exist; find --json marks each with "coverage": "catalog_only" in its metadata, and search never opens them.