Inference data and retention

Updated 2026-11-13. Eldrim Compute is operated by Happi Hacking AB, org. nr 556912-2707, Sweden.

For marketplaces and integrators

Short answers for marketplace listings and integration reviews, as the service runs today. Each answer links to the text that explains it.

QuestionAnswer today
Training on customer dataNo. We do not use prompts, responses or other inference content to train models. Berget AI AB runs some of the models we offer. Berget’s published terms say it does not store prompt or output content; they do not address training (what happens at the gateway, execution providers).
Prompts and responses in our databaseNo prompt or response columns. The database stores usage and routing metadata. Records written before our release of 8 October 2026 keep the raw error text of failed routing attempts, and we have not established that those records are free of content (what happens at the gateway, what we retain).
Prompts and responses in gateway logsKept out by design. Content-bearing errors are logged as error categories, and crash dumps are blocked. Diagnostic logs rotate by size, with no fixed period in days. We have not established that logs and backups from earlier software are free of content (what happens at the gateway, what we retain).
Retention of usage metadata and billing recordsAs the privacy policy sets out, billing records and the usage metadata they are based on are kept for 7 years, as Swedish bookkeeping law requires. Response times and routing details are kept with those records until a separate period is set. Our deletion and backup-expiry procedures are under review (what we retain, privacy policy).
Zero data retentionYes for the models Berget runs, on Berget’s published commitment never to store prompt or output content; our gateway does not store it either. No for our test capacity, whose hosts can page memory to disk. Each response names the provider that answered (current assurance, execution providers, prefix caches).
Where requests are processedSweden. Every route we have configured is set to Sweden (se), and requests and failover stay within your account’s residency policy. Our gateway runs on a server we rent in Sweden, and our test capacity on machines we run in Sweden. For Berget, the setting rests on Berget’s own statement that its inference runs in Stockholm. Berget’s published data processing agreement says its data centres are within the EEA. Whether that agreement covers our customers’ data, and a written commitment to Swedish processing, are still open (execution providers, residency).
Capacity for marketplace accountsNone at the moment. Our dedicated GPU server in Sweden is out of service while its rental agreement is renewed; this row changes when it returns. Berget’s models are closed to accounts we set up for a marketplace. Our test capacity opens only to our internal account and to accounts we mark as testers or coworkers, and the gateway has no separate block that keeps marketplace accounts off it. Further capacity for marketplace accounts gets its own host, logging and provider review before it serves them (execution providers, prefix caches).
Sub-processors that handle request contentBerget AI AB, Stockholm, for the models it runs, and the hosting company in Sweden that rents us our gateway’s server. Both are in our list of processors. Berget lists its own subprocessors in its data processing agreement. Accounts we set up for a marketplace carry a marketplace setting, and the gateway does not route such accounts to Berget (privacy policy).
Residency recordEach response from a provider carries an x-eldrim-residency header with the jurisdiction configured for the route that served it. When the request is recorded, its request record from GET /v1/eldrim/requests/{id}, under the same x-request-id, holds the same value. Both report configuration; where requests are processed says what that rests on (residency).
Contactsupport@eldrim.ai, with a request ID and a synthetic reproduction. For an enterprise review, ask us for the exact routes, provider commitments and evidence for your workload (support). Happi Hacking AB, Sweden, operates Eldrim Compute.

Current assurance

Eldrim’s gateway processes inference content without deliberately storing prompts or responses in its application logs or database. Its request handlers, failure logs and crash-dump controls are tested with synthetic content.

We do not currently offer a verified, end-to-end zero data retention guarantee across all routes. Our test hosts use swap or operating-system paging, which can write memory to disk. Berget’s content-retention commitment rests on its published terms. Workload placement and provider terms need review before you send sensitive data.

Zero data retention with Berget. Section 7 of Berget’s published terms says it will never store prompt content or generated output, and section 6.6 keeps customer data out of its backups. Our gateway does not retain that content either. For these models the answer rests on Berget’s commitment, not on an audit by us.

What happens at the gateway

Prompts, responses, tool arguments, embeddings input and System One state pass through memory to the selected execution provider and back to your application. We do not use that content to train models. The database stores usage and routing metadata, with no prompt or response columns. Content-bearing exception terms are reduced to error categories before being logged or recorded as route failures.

The gateway blocks Erlang crash dumps and native process core dumps. Its host had no active swap at the check on 8 October 2026. These controls concern the gateway; the provider running a model has its own memory, logging and storage configuration.

The Responses endpoint is stateless: stored responses, background jobs and server-side conversation continuation are unavailable. Our API does not provide hosted file or conversation storage. Your application and coding tools may keep their own history.

What we retain

Operational records include account and key identifiers, request identifiers, model and provider, execution jurisdiction, timestamps, status and classified failures, token counts, latency, prices and costs. Account, billing, security and support records are separate from inference content. Their purposes and retention are described in the privacy policy; the operational deletion and backup-expiry procedures remain under review.

Gateway diagnostic logs rotate by size. Rotation is not a promise that every record expires after a fixed number of days. Database backups contain the metadata and account records in the database. This release does not establish that all historical logs or backups are free of content from earlier software.

Execution providers

Berget. Section 7 of Berget’s published terms says it does not store prompt or generated output content. That is a provider commitment, not an infrastructure audit by Eldrim. Confirmation of the downstream processing agreement and contractual Swedish processing remains open.

Our test capacity. Test models run on our own workstations. Their operating systems can page memory to disk, and their logging, crash collection and cache configuration have not passed an end-to-end retention review. They are outside a verified zero-retention assurance. Invitation and model-access restrictions still apply.

Our dedicated capacity. Our rented GPU server in Sweden, which passed our host review of 13 October 2026, is out of service while its rental agreement is renewed. No request reaches it until it returns, and this section will say when it does.

New capacity. Dedicated or rented capacity needs its own recorded host, logging and provider review before any stronger assurance applies. A Swedish location alone does not establish zero retention.

Memory and prefix caches

Serving engines may retain token-derived key/value tensors in memory to reuse a prompt prefix. This can outlast an individual request. It is not model training, and it is not a guarantee of immediate memory erasure. Disk-backed cache offload, swap, crash collection and cache isolation all matter. We have not verified every model host against those conditions. On the model servers we run with vLLM, the gateway gives each account its own cache key, so one account’s requests never reuse another account’s cached prefix.

Support and binding legal obligations

Contact support@eldrim.ai with a request ID and a synthetic reproduction. Content you choose to send in a support message becomes support correspondence and falls outside the inference-content handling described above.

This statement does not override binding legal obligations. Any legally required preservation would need a separate assessment, with notice where permitted. We cannot recover inference content we have not retained.

For an enterprise review, ask us to identify the exact routes, provider commitments and operational evidence for your workload. This statement is not a certification or a substitute for the applicable contract.

Developer guide · Privacy policy · Terms