Skip to main content

​
Excluding automatic saved memories

Reliant normally attaches saved global, project and repository memories to every model turn, including a new thread. Set exclude_ambient_memory: true to omit those automatic memory blocks:
Or set it on a workflow, loop or router thread boundary:
Omitting the flag preserves existing behavior. An excluded ancestor cannot be re-enabled with false or an omitted flag. Exclusion is monotonic and durable for a thread: after its first excluded call (or excluded child initialization), subsequent turns, retries, coarse restarts and stopped-agent resumes retain it. A new/fork child also inherits excluded parent lineage. Use a separate fresh thread outside that lineage for work that needs the default memory behavior. The additive database field defaults to false for existing threads; it cannot prove what older runtimes already sent, nor retroactively remove prior content. Policy read/write failures stop the call rather than send memories by default. The lineage must reach a real root: missing ancestors, cycles, or ancestry beyond 256 threads stop the call before any provider request, rather than treating unresolved policy as false. Repair that lineage before retrying. This is not full context isolation or an approved-information-only mode. It only excludes automatic saved-memory injection; it does not erase history or restrict tool permissions. A fork/inherited thread can still contain private messages, summaries, tool results and attachments. Explicit prompts and skills can contain private information too.

​
What still reaches the model

  • Reliant’s identity/safety bootstrap and static capability routing guide.
  • Available working-directory/worktree and multi-repository names/paths.
  • The available skill catalog (names and descriptions), even with skills: [].
  • Explicit system_prompt, messages, thread injects, preloaded skill bodies and any preset content that supplied those values. Forge start-here may be added when preloading skills in a Forge project.
  • Conversation history, compaction summaries, attachments and tool results; queued user/sub-agent messages delivered before history is loaded.
  • Offered tool names, descriptions and schemas, including project-specific MCP, integration and spawn/preset descriptions when those tools are offered.
  • Machine-less and bug-report guidance when applicable; sandbox runtime guidance is retained even though it historically shared the saved-memory block.
  • Provider-specific guidance and transport metadata. For example Claude Code’s base prompt, OpenAI-family tool guidance and Antigravity’s managed prompt are independent of saved memories.
Empty preloaded_tools and loadable_tools suppress the tool array and future dynamic tool loading. Empty skills suppresses skill bodies, not the catalog announcement. These declarations do not suppress the system bootstrap. A machine-less run removes machine access, not model context. A project-less launch is not currently supported (project_id is required); an empty project configuration is not the same thing as no project. Captured-request tests verify that machine-less + empty tools/skills still carries bootstrap instructions and, when configured, the skill catalog.

​
Before enabling an approved-only application

Independently inspect the complete serialized request for the intended provider and deployed runtime: system/bootstrap/provider instructions, history and summaries, attachments, explicit/preset prompts and skills, catalog metadata, tool schemas/results, queued messages, paths/repo metadata and transport fields. Exercise nested execution, retries and recovery, not just a memory formatter. Use synthetic private canaries; do not send real private data as a test. The Growth AI path remains disabled until that full boundary is proved and the runtime supporting this flag is deployed. This change does not deploy a daemon, release Reliant, or authorize enabling Growth AI.