Cursor

Hansard reads the Cursor CLI’s own conversation stores, the Cursor app’s agent transcripts, and the app’s older SQLite history. Cursor cloud agents are a separate source that you import on request, and they always stay separate from local Cursor sessions.

Source IDs

Source	Use it for
`cursor-cli`	Cursor CLI (`cursor-agent`) conversations: interactive and headless `-p` runs from `~/.cursor/chats` and ACP sessions from `~/.cursor/acp-sessions`, including saved Cursor CLI account profiles.
`cursor`	Cursor app agent transcripts and older local SQLite history.
`cursor-cloud`	Cursor cloud agents, read from Cursor’s API. Runs when you ask for it or turn it on in Settings → Imports → Cloud agents.

Stats, filters, and the session header show where each conversation ran: Cursor CLI, Cursor cloud agents, or the Cursor app release it came from (Cursor 1, Cursor 2, or Cursor 3). The app keeps every release’s conversations in the same stores and records no version for each one, so Hansard assigns the release from the conversation’s start date (2.0 from October 29, 2025, 3.0 from April 2, 2026; earlier releases, 0.x included, count as Cursor 1), lowered to the app version Hansard recorded when it imported the conversation. Filter with hansard history --provider cursor-cli, cursor-3, cursor-2, or cursor-1.

Recovered Cursor app conversations use the same generation rule. The grouping refreshes without a reimport. A recovery record without a usable start date still has an unknown generation.

How the import works

For the Cursor CLI, Hansard reads each conversation’s store.db read-only from a copy, so the running CLI is never disturbed. The store keeps the model behind every reply (including the model Auto picked), tool calls with their results, per-step timestamps, the branch and time zone, and how full the context window was. Headless -p runs count as scripted jobs, like other SDK sessions, while interactive and ACP conversations count as your own. When the CLI also writes an agent transcript for a conversation, Hansard archives the richer store and replaces a transcript-only copy imported earlier once the store’s copy holds everything that copy recorded. A simulated prompt recorded as provider-generated system context can hold a transcript twin’s matching user prompt. A prompt or reply that cannot be compared keeps its copy. Saved Cursor CLI account profiles are read from their own folders.

For the Cursor app’s current history, Hansard scans Cursor’s agent transcript folders for each project. It groups files by session and reads messages, tool calls and results, errors, models, usage, and subagent links. The project comes from the recorded path or from the project folder name. For older history, Hansard also reads Cursor’s workspace and global SQLite databases, including Composer, chat, and message records, and puts them in the most complete order the database allows. A session that has only prompts or applied edits left imports as a partial conversation.

Cursor can leave the same Composer session in several places. Hansard combines them only when they share the exact Composer ID. The agent transcript takes priority over the global database, which takes priority over workspace data and partial recovery. Lower-priority copies can still add a title, working directory, timestamps, usage, or the original files. Sessions with similar text but different IDs stay separate, so one conversation does not appear several times and unrelated work is never merged.

The watcher follows transcript files as they change. A full --since all import also recovers older conversations from the SQLite databases, so run it first after you install Hansard. The separate cursor-cloud command reads cloud agent details, usage, and any message text Cursor’s API returns. Cloud agent IDs and local Composer IDs never overlap, so a cloud agent always stays its own conversation.

The cloud import also reads each retained v1 run stream. It preserves thinking, tool calls, JSON arguments, JSON results, and truncation flags. The legacy v0 conversation and run streams have independent ordering, so stream activity appears as a separate block. Hansard does not invent per-message clocks or model identities. SDK interaction updates remain raw evidence, including any turn usage or nested tool details that the simplified events omit.

Run streams expire. An expired stream retains its archived events and terminal run metadata, but Hansard cannot recover events Cursor no longer returns. Large tool values can also be absent with an explicit truncation flag. Cloud attachment references stay raw and do not establish a downloaded file.

hansard init and hansard doctor count older SQLite entries without decoding them, so the number they show is an estimate; the import finds the exact count. Hansard applies --since before it reads message rows, reads large databases in batches, and prints progress as it goes. A --since all import can still take a while when Cursor’s state.vscdb is several gigabytes, because every selected conversation is decoded once.

hansard import --source cursor --since all
hansard import --source cursor-cli --since all
hansard import cursor-cloud --since all
hansard index rebuild

cursor-cloud uses CURSOR_API_KEY or cursor.apiKey in Hansard’s config. It only reads from Cursor, and Hansard never stores the key in the archive.

With Automatic source selection, the watcher already includes cursor and cursor-cli. With a Custom list, enable them. The watcher picks them up within 30 seconds:

hansard config sources enable cursor
hansard config sources enable cursor-cli

To keep cloud agents current, turn on Cursor cloud agents in Settings → Imports → Cloud agents. The watcher then checks it every minute, after any earlier scheduled Cursor cloud import finishes. Capture remains subject to API retention, rate limits, and import queue delays. Otherwise, run the command again when you want newer cloud agents.

Recover models and token counts

Most local Cursor conversations record no model or token counts. Cursor keeps a usage history for each account that names the model, the tokens, and the conversation of every request. To add it to your archive:

hansard import cursor --recover

Or open Settings, then Recover Cursor models and usage.

It reads every account signed in on this computer: the Cursor app, the Cursor CLI, and saved Cursor CLI profiles, each account once. It only reads, and Hansard never stores or prints a token.
Each request joins the local conversation with the same Composer id. Usage never creates a conversation.
Cursor keeps a limited history. Hansard keeps every request it has recovered, so later imports retain them after they leave Cursor’s history. A conversation’s totals count as exact only when it began inside that history.
Conversations made under an account that is not signed in on this computer need Cursor or the Cursor CLI signed in to that account. Sign in, run the command again, then switch back.
Cursor CLI conversations take their token counts from the same history, so run it after CLI work to fill them in.

Review the import in the app

When the import finishes, open Sessions and selectFilter & sort.
Set Source to Cursor or Cursor cloud agents and setTime to the window you imported. UseFolder to narrow the list when you use this harness in several projects.
Select a conversation to check its messages, tool calls, files, commits, model, and token usage.

To keep archiving new sessions from a local harness, openSettings → Imports. Automaticfollows every supported harness installed on this computer.Custom lets you turn individual harnesses on or off. Cloud sources turn on separately under Cloud agents, and ChatGPT, Claude.ai, Gemini, and Gemini Notebook / NotebookLM exports import withImport export…. Each also has its own import command. If a conversation you expect is missing after you reset the filters, run the import command on this page again with --explain-skips where the source supports it.

What Hansard preserves

Current agent transcripts with tool calls, errors, model settings, and workspace details when Cursor wrote them.
Context window snapshots and Cursor’s cost buckets from older SQLite history, kept separate from input and output tokens.
The model on each reply in the older chat panel, and the models older Composer conversations billed.
Cloud agent status, run times, duration, branch, pull request, token totals, available message text, and retained stream tools and thinking.
Both the model and the harness, so Composer appears in model views and Cursor appears in harness views.

Recovery boundaries

An older app conversation can have only prompts or edits left in its local database. Opening that conversation in Cursor and sending a short prompt may make Cursor write a modern transcript. This action changes native history and can use model quota. Hansard does not perform it automatically, and the route still requires native validation. Usage recovery with --recover is a separate read-only operation. It cannot recreate missing conversation text.

Raw replay through hansard rebuild currently supports modern Cursor agent transcripts. Cursor CLI stores, app SQLite stores, and cloud captures preserve raw evidence but do not have a built-in replay entrypoint. Repeat their import to refresh parser output. Repeat hansard import cursor-cloud --since all to read available stream evidence for older cloud captures. Expired events remain unavailable.

Known limitations

promptTokenBreakdown, contextTokensUsed, and contextTokenLimit show how full the context window was, not billed usage.
Composer conversations from before Cursor saved a model setting with each conversation record a model only when Max mode billed it, and only for the conversation as a whole. Their replies carry no model, and other conversations from that period carry none unless recovery finds them in Cursor’s usage history.
Simplified cloud run streams supply no model ID or per-event clocks. Expired streams cannot supply missing tools. Available conversation text, earlier stream evidence, and terminal run replies remain usable.
Local Composer IDs and cloud agent IDs never overlap, so Hansard never treats a local session and a cloud agent as the same session.
Cursor CLI stores retain context-window counts, but the native sample has no token-usage totals or CLI version. Usage totals come from recovery.
Cursor 1, 2, and 3 follow release dates, so a conversation started on an older release after a newer one shipped counts as the newer release unless Hansard recorded the older version when it imported it.

Refresh an existing archive

hansard import --source cursor --since all
hansard index rebuild

A full reimport is the clean way to update Cursor history that an older Hansard version imported.