Skip to main content

​
Immutable Decision Trace Store

The OSS explain() only works within the active session and captures enriched trace data (query events, error events, session timing, state diffs). Pro builds on this with persistent, cryptographically signed, immutable traces — the full causal chain is queryable forever.

​
OSS Trace Enrichment

Every commit_outcome automatically captures:
  • Session timing — session_started_at, session_ended_at, session_duration_ms
  • Query events — every search() and list() call with parameters, result counts, and per-operation latency
  • Error events — any errors during reads, writes, or tool calls
  • State diff — entries created, updated, and confidence changes during the session
  • Full entry snapshots — value, memory_type, and written_by captured at read time

​
Pro Immutable Traces

Pro adds:
  • Cryptographic integrity — HMAC-SHA256 content hashing with Merkle chain linking across sessions
  • LLM call spans — model, provider, prompt/completion token counts, cost, latency, temperature, finish reason
  • Write events, tool calls, agent interactions — full record of every action
  • Token and cost aggregates — total_llm_calls, total_tokens, total_cost_usd per trace
from amfs_traces import TraceRecorder, InMemoryTraceStore

recorder = TraceRecorder(memory, store, account_id=acct.id)

# Reads and external contexts are tracked automatically
recorder.memory.read("svc", "retry-pattern")
recorder.memory.record_context("pagerduty", "3 SEV-1", source="PagerDuty API")

# Record LLM calls for token/cost tracking
recorder.record_llm_call(
    model="gpt-4o", provider="openai",
    prompt_tokens=1200, completion_tokens=450,
    latency_ms=2340, cost_usd=0.0285,
)

# Outcome commits automatically persist the trace
updated, trace = recorder.commit_outcome("DEP-500", OutcomeType.SUCCESS)

# Months later, explain still works
result = recorder.explain("DEP-500")

# Search across all decisions
traces = recorder.search_traces(entity_path="checkout-service", outcome_type="critical_failure")

​
OpenTelemetry Export

Export AMFS decision traces as OpenTelemetry spans for integration with existing observability stacks (Jaeger, Grafana Tempo, Datadog, Honeycomb). Traces are mapped to hierarchical spans following GenAI semantic conventions — each LLM call becomes a gen_ai span, and memory operations become amfs spans. 
from amfs_otel import TraceExporter, configure_otel

configure_otel(endpoint="http://localhost:4317", service_name="my-agent")
exporter = TraceExporter()
exporter.export_trace(trace)  # sends spans to your OTel collector