Claude Science

Claude Science keeps a separate local database. It does not use Claude Code’s conversation store. Hansard reads Science frame messages and artifact revisions from ~/.claude-science/operon-cli.db, each ~/.claude-science/orgs/<organization>/operon-cli.db, and their adjacent artifact directories. The same stores support the desktop app and its local browser UI.

Import local history

hansard import --source claude-science --since all

Enable Claude Science in Settings → Imports for ongoing capture. HANSARD_CLAUDE_SCIENCE_ROOTS selects custom data directories as a path-separated list. Each custom directory can contain a database directly or organization stores under orgs/. Database and write-ahead-log changes trigger the source watcher, including databases created in new organization folders.

Preserved data

Data	Capture behavior
Frame messages	Text, thinking, tool arguments, results, errors, and embedded attachments, including the contents of text document attachments.
Structured results	Native failures, partial artifact saves, server results, citations, and tool references. Unknown blocks remain readable.
Background execution	Completion notifications match the original tool and execution identities. Pending output stays linked to the call that started it.
Full output and images	Referenced local output files, disk images, and cell-image sidecars, with size-limited reads and hash checks.
Compaction	Retained messages join the active history by message identity. Documented compaction changes preserve earlier content. Conflicting revisions stop the update.
Artifact revisions	Available local bytes, recorded hashes, version identity, and metadata.
Harness notices	Skill suggestions, compute state, plan approvals, and output notices remain expandable context. Native tool attachment receipts do not create user turns.
Parent and child frames	Separate conversations with explicit links, agent identity, task descriptions, native message and tool counts, and size-limited prompt/result previews.
Models and usage	Recorded reply models, native message usage, and frame counters. Cached input is counted once. Selected models stay configuration. Missing values remain unknown.
Auxiliary history	Execution logs, annotations, notes, events, reviewers, routines, and compute records appear as source context. Alternate branches, system prompts, reproducibility snapshots, and scoped memories remain raw evidence.
Account and project memories	The memory backup preserves all permitted memory records as read-only JSON snapshots, including records without frame links. Native scope, origin, supersession, and revision history remain available.

The importer reads only allowed history tables and columns. It excludes credential tables and never archives the complete database. Missing per-message clocks use stable session ordering; native message clocks are preserved when present. Unknown blocks remain in raw evidence and produce visible accounting. Unknown database columns are excluded and counted.

Recovery and limits

hansard rebuild --from-raw --source-type claude-science-history

Offline replay verifies the imported history, artifact bytes, and tool files. Captured child summary evidence also replays without opening the provider store. Source restore does not write the imported JSON over the Science database. A stale capture that loses messages, background results, or captured bytes leaves the earlier archive intact and reports an import error.

Child summaries retain up to 1,600 characters for each prompt and result preview. The linked child conversation retains its full captured transcript. Selected child models remain configuration, separate from models recorded on assistant messages.

The importer processes one frame at a time and retains only identity fingerprints between frames. The reader limits each frame to 100,000 history rows, its imported history to 128 MiB, combined file bytes to 128 MiB, and each file to 32 MiB. Remote-only files and files outside the configured store remain references. Alternate branches retain raw history without merging their turns into the active conversation.

The observed local format is Science 0.1.56 on macOS. Bundled examples establish schema and artifact evidence. Native checks exercise tool failures, artifact revisions, background completion, disk images, and large outputs. Synthetic checks cover server tools, unknown blocks, cancellation, and unsafe paths. A resumed PubMed connector call and compute inventory reached the archive automatically and appeared in installed Hansard with their tool results and reviewer link. All 13 observed sessions passed repeat import and offline replay. The native memory trial captured three records. Synthetic fixtures cover records without frame links, metadata changes, and separate organizations. Other releases and platforms, individual connectors, and remote compute jobs need separate checks. The inspected client has no enabled routine interface, and the inspected account has no configured compute provider.

hansard memory backup captures Science memories separately from conversations. The supervisor repeats the backup every 15 minutes. Removed provider records leave archived memory revisions intact. Manage live memories in Claude Science.

Rerun the import command after upgrading to acquire expanded tool evidence, corrected token totals, child summaries, and the contents of text document attachments that earlier imports kept without their bytes.