--- description: Understand local storage, provider choices, agent context and encrypted cloud boundaries. --- # Privacy and providers Your memory source is readable Markdown on your disk. You choose which agents use it, whether extraction uses a provider, and whether to keep encrypted snapshots remotely. Those choices have different data boundaries. ## What leaves your machine | Feature | Data boundary | Default / control | | --- | --- | --- | | Local memory storage | Source files and indexes stay in the configured local directory. Anyone/process with filesystem access may be able to read them. | Local files; use your operating system's access controls and disk protection. | | Local embeddings and reranker | Memory/query processing is local; first use may download model weights. | `embedding_provider=local`; reranker enabled. | | Ollama provider | Text goes to the configured Ollama endpoint. “Local” depends on where that endpoint runs. | Opt-in provider/model/base URL. | | LiteLLM generation provider | Selected conversation/file text and maintenance inputs can be sent to the configured model provider. | `llm_provider=none` by default; opt in deliberately. | | Remote embeddings | Memory text and queries go to the selected embedding provider. | Opt-in `embedding_provider=litellm` or a remote Ollama endpoint. | | Connected coding agent | Recalled or whispered memory becomes part of the agent's context and follows that agent/provider's processing policy. | Depends on the client you connect and use. | | Encrypted cloud snapshots | The client encrypts locally; remote object storage receives encrypted bytes. The service handles relevant account, device/store, snapshot and operational metadata. | Cloud backup disabled until configured; entitlement and recovery readiness also apply. | | Local diagnostics | Logs/database records can contain prompts, memories, candidate traces and feedback. They are local operational data, not part of a source-memory snapshot. | Review/redact before sharing diagnostics. | ## Choose an extraction mode - **No separate Ormah LLM:** explicit memory tools and local retrieval work. Raw transcript/file extraction is unavailable unless a caller provides pre-extracted memories. - **Ollama:** configure an endpoint and a model you have made available there. Restart Ormah after configuration changes. - **LiteLLM:** install the optional extra, choose a model, and explicitly configure credential selection. `llm_api_key_env_var` names the environment variable containing the key; it is not the key itself. `llm_inherit_api_key=false` prevents ambient-key inheritance by default. Choosing agent-driven maintenance can use the connected agent to classify maintenance work. It does not supply the server with a transcript-extraction provider. [Setup](/docs/getting-started/setup) explains the controls. ## Local service access Ormah defaults to `127.0.0.1:8787`. General agent/admin routes are designed for a local trusted environment; they are not all protected by an account login. CORS allows loopback browser origins, which is different from authenticating every API caller. Sensitive account/protection routes additionally require a loopback caller and a local capability header. That capability is separate from your cloud account credential. Do not publish either. Changing the bind address or forwarding the port does not create a securely authenticated remote memory service. See [HTTP API](/docs/reference/api-surface#authentication-and-errors). ## Ownership and portability You can read and move source files, but a snapshot is not a complete copy of all runtime state. Configuration, keys, operational feedback and logs need separate handling. Use [Back up and recover](/docs/use-memory/backup-and-recovery) for the supported local path and the detailed cloud reference. Review session/file capture directories before enabling them. Exclude sensitive files you do not want extracted, and inspect the resulting memories. Deleting a memory from Ormah cannot remove copies already delivered to another agent or sent to a provider.