SOMA Adaptive Memory Intelligence™ (SOMA AMI™) — Frequently Asked Questions

What is SOMA Adaptive Memory Intelligence™ (SOMA AMI™)?

SOMA AMI™ is a deterministic decision plane and memory engine for AI agents. It structures state into ATOMs (typed facts), applies policy-driven arbitration over that state, and governs which tools, workflows, or agents may run, with a full, auditable decision trace.

Who is SOMA AMI™ for?

Teams building LLM/agent workflows that must be repeatable, auditable, and safe—customer support, humanoid robots, RevOps, risk & trust, RPA/automation, and internal agent platforms.

What problems does it solve?

SOMA AMI™ addresses critical challenges in AI agent systems by providing a centralized, deterministic, auditable decision plane for agent decision-making. When agents produce flaky or non-repeatable decisions, SOMA AMI™ ensures consistent outcomes through deterministic selection algorithms with built-in tie-breaking mechanisms. Multi-agent conflicts are resolved through sophisticated arbitration using top_k guards that prevent conflicting actions from executing simultaneously.

What are ATOMs?

ATOMs: Adoption Translation and Observation Model — the canonical schema for memory facts. They standardize how agents express state, tags, and observations so downstream decisions are consistent. ATOMs are typed facts that serve as the only inputs to SOMA AMI™.

What is an EMU?

An EMU is an Executable Memory Unit. It has a trigger, action and policies associated with it.

How does SOMA AMI™ guarantee determinism?

Each decision is a pure function of (atoms, registry_version, policy_version, ts_bucket, events_read_cutoff_ts). The trace records all five so replays are identical.

What is top_k and what are the guardrails?

Default top_k=1. If top_k>1, winners must be advisory and in disjoint exclusion groups; ungrouped EMUs are ineligible. If a non-advisory candidate would pass, AMI returns 400.

How does recent_event(...) work?

It matches normalized subject, verb, object within (context_ts − duration, context_ts], considering only persisted events at or before events_read_cutoff_ts plus any context EventAtoms.

What is the Trigger DSL?

A compact, boolean expression language that evaluates state conditions, tags, and recent events to determine when actions should be triggered. The DSL supports logical operators (AND, OR, NOT) and functions for state evaluation, tag matching, and temporal event queries within specified time windows.

What are shadow and dry-run modes?

Shadow returns selected: [] and marks would-be winners in the trace; dry_run computes selection but skips cooldown/idempotency side-effects.

What is in the decision trace?

A comprehensive record including the registry/policy versions, evaluation context, inputs summary, every candidate with pass/fail reasons, suppression codes, scores, and the final selected results.

What are the EMU lifecycle states?

draft (registration default), shadow, canary, active, inactive, archived. Transitions are enforced by the lifecycle API; soft deletes only (you can read history).

How do registry/policy version pins work?

Use options.pin_registry_version and options.pin_policy_version. If they do not match the server's active versions, AMI returns 412 VERSION_MISMATCH.

Which regions are supported?

us-east-1, eu-west-1, ap-southeast-1. Choose per operator in config.

What data does SOMA store and for how long?

SOMA stores events, decision traces, and operational metadata with configurable retention policies to ensure compliance while maintaining system performance and auditability.

Does SOMA AMI™ execute actions?

SOMA AMI™ returns a payload (e.g., tool_call) for selected EMUs; your host executes the action and emits events back to the system.

How does SOMA AMI™ handle ML-derived facts?

Atoms can include source:"ml". If allow_auto_with_ml_atoms=false, auto EMUs depending on ML facts are suppressed with POLICY/policy_auto_ml_blocked.

Does SOMA support no-PII configurations?

Yes. Register ID-only EMUs (no action); /ami/invoke returns the EMU id/intent/policy without payload. You can map EMU ids to actions inside your secure environment.

Glossary

• Atom / ATOM — typed fact; abbreviation of "Adoption Translation and Observation Model."
• EMU — registered decision unit (trigger + policy + optional action), versioned with lifecycle.
• Trace — full, deterministic rationale for each decision.
• Cutoff — events read cutoff timestamp that bounds persisted events used in evaluation.

Ready to Build with SOMA AMI™?

Get access to our complete Developer Documentation, API references, integration guides, and technical support to start building deterministic AI agent workflows.