What Fungi keeps

Fungi saves conversations, files and work records so Agents can continue their work. Where those records live depends on where the Agent runs.

Local storage does not mean that nothing leaves your machine. Prompts, conversation context, files an Agent reads and tool results can be sent to the selected model provider. Connected services and remote tools can receive data too.

We have not verified any Fungi setup in which Fungi, model providers and connected services keep no copies of your data. Using your own key or subscription, turning off analytics, or choosing a provider’s no-training option does not delete Fungi’s saved records.

The Local and standalone Oyster behavior below is checked against Fungi 0.6.0. An operator’s configuration, backups and provider terms are separate from what the application code stores.

Local: on your machine

Local is the Agent runtime used by the terminal. Attaching your machine as a Computer for a hosted Agent does not make that hosted Agent’s conversation local.

What	Where it lives	How long it remains
Sessions and transcripts	An Agent’s SQLite database in project .fungi/ or home ~/.fungi/ state	There is no general timed Session expiry. Compacting the context sent to a model does not erase the saved history.
Files and build outputs	The working directory you give the Agent	Until removed by a file operation. Stopping or archiving an Agent does not delete its project files.
History saved when a conversation is shortened for the model	Separate archive segments beside the Agent database, when the built-in archive writer is used	These can remain after a Session is deleted. An installed extension can supply its own archive handoff.
Background jobs and messages	The Agent’s durable work stores, including Watchdog jobs	Finished jobs can keep their inputs and results. Deleting a Session does not remove every copy of its content from these other records.
Credentials	By default, ~/.config/fungi/credentials.json	Until removed or replaced. This includes provider keys, subscription tokens and device private keys. File permissions restrict access; they are not an encryption-at-rest guarantee.
Error logs and usage records	Home logs and project .fungi/processes/ state	Service logs rotate across four files of up to 1 MiB each; terminal logs retain the current file and one rotated file. Rotation is by size, not days. Command output can remain after a command finishes or is stopped. It is removed when Fungi cleans up the associated process records.
Temporary files used while running commands	$XDG_RUNTIME_DIR/fungi/tmp or ~/.cache/fungi/tmp	Cleanup runs on normal close and supported error paths. A crash or interrupted process can leave files behind.

Deleting a Local Session removes its active conversation records and cleans up its background processes. It does not securely erase the database file or delete separate history archives, project files, background-job records or your backups.

Archiving a Local Agent hides it and prevents ordinary use; it does not move, export or delete its stored data. This also applies to Local Agent archives created with Fungi 0.5.0. Signing out removes selected credentials, not your conversations or files. Remote sign-out cleanup can remain pending when a service is unavailable.

Fungi’s local state also includes settings, installed extensions, skills, caches and staged runtime binaries. Deleting one Session or signing out is not a whole-installation reset.

When Local is upgraded, older Agent and extension databases may be moved to ~/.fungi/archive/ if the new version cannot use them. Deleting a current Session does not delete these older databases. Keeping them does not guarantee that the current version can open them.

Oyster: in your operator’s storage

Oyster is a self-hosted installation. Its operator chooses the storage provider, logging and backup arrangements.

What	Where it lives	How long it remains
Sessions, transcripts and durable work	The Agent’s Durable Object storage	There is no general automatic deletion schedule for saved Sessions. Checks that limit new work when storage is full do not delete existing conversations.
Computer files	The Computer’s durable Workspace storage	Files survive runtime sleep. An idle process stopping is not Workspace erasure.
Apps, build records and artifacts	App databases, installation records and separately configured artifact storage	Deleting an App may leave its stored build files and records showing that it was installed or removed. A build timing out does not delete its record.
Credentials and access records	Operator configuration and the relevant Agent or credential store	Revoking a named key can leave its hash and metadata. Removing a connection is different from deleting every copy of its credentials.
Logs, archives and backups	The operator’s configured storage and services	There is no single verified retention period covering them all. Ask the operator what is enabled and how deletion is checked.

Oyster deletes some older installation-history entries after 90 days, once they have been replaced and no longer need recovery. It keeps the latest entry and entries still needed for recovery. This is not a 90-day deletion rule for conversations, files, builds or every installation record.

Deleting an App’s database, revoking an access key or removing a Worker does not establish that separate artifact buckets, backups or provider copies have been erased.

Hub: Fungi’s hosted service

Hosted Agents and Computers use server-side storage even when you open them from your own terminal or attach a local folder.

What	Where it lives	Retention and deletion limits
Sessions and transcripts	Agent storage on the service	They survive process restarts. There is no verified general automatic expiry for every saved Session. Compaction does not mean deletion.
Computer workspaces	The Computer’s persistent storage	Suspension retains the Workspace. Removing a runtime is not proof that every storage-provider copy has disappeared.
Apps, builds, releases and generated media	App state, release catalogs and separate artifact or media stores	These have separate lifecycles. Deleting a conversation or uninstalling an App does not establish all-copy deletion. Some release and erasure metadata is deliberately retained.
Account, sign-in, Team and billing records	Account and product databases and their service providers	A sign-in token expiring does not establish that its database row, IP address or browser details were deleted. Closing a Team is not erasure.
Credentials	The relevant protected credential stores	GitHub sign-in tokens are configured for encrypted storage. This does not mean every file, log or user-supplied secret is encrypted or absent from a Computer.
Error logs and usage records	Work stores, service logs and any configured log or telemetry receiver	Settled work can retain payloads. Disabling one analytics or telemetry feature does not mean there are no logs. Receiver retention is separate.
Optional archives and backups	Separately configured archive and backup storage	Deleting live records does not automatically erase these copies. There is no verified 30-day upper bound covering every archive, log and backup.

There is no button to delete or export your whole account. Contact Privacy and support to ask what can be removed or exported, what must remain, and how completion will be confirmed. Deleting data from one place does not mean every copy has been deleted.

The Hub cannot erase copies on your machine, in your own backups or in a service to which you separately sent them.

Model providers keep their own records

Provider policies depend on the route, account, feature and settings you use. No training is not the same as no retention, and a personal subscription does not automatically receive API or commercial data terms.

• Claude subscriptions: Anthropic’s consumer policies cover Free, Pro and Max, including those accounts in Claude Code. Deleted chats normally leave backend storage within 30 days. Data opted into model improvement may remain de-identified for up to five years; feedback, safety and legal exceptions can retain data longer. These are not the same rules as its commercial service. See consumer retention and consumer training choices.
• Anthropic API and commercial products: the standard API retention policy deletes inputs and outputs within 30 days, with feature, agreement, covered-model safety and legal exceptions. Commercial products with saved conversations retain those separately. No-training defaults do not remove retention. See commercial retention and commercial data use.
• ChatGPT subscriptions: do not apply OpenAI’s API retention or no-training rules to your subscription. Check the account’s consumer data controls and the terms for the feature you use.
• OpenAI API: abuse-monitoring logs and feature-specific application state are different. The usual up-to-30-day abuse-monitoring period is not a universal deletion deadline. Zero Data Retention requires approval and has feature limits; it is not an end-to-end Fungi setting. See OpenAI API data controls.

Other text, image and video providers, gateways and connected services have their own terms. Do not assume an unverified provider has no retention. Fungi’s Privacy Policy and Subprocessors describe the hosted service’s disclosures.

Before sending sensitive data or requesting deletion

1. Check where the Agent runs, which Computer it uses and which provider receives the request.
2. Do not paste credentials into prompts. A credential stored in the right settings screen is different from a key copied into a file or conversation.
3. Keep your own backup of important work, and remember that this creates another copy to manage.
4. Treat archive, compact, sleep, sign out and revoke as distinct operations, not synonyms for delete.
5. For deletion, name the Sessions, files, Team or account involved. Ask separately about archives, artifacts, work records, logs, backups and provider copies. Do not assume every copy will be deleted immediately or made impossible to recover.