# Sibyl Phase A Product Memory Suite

Run: `run-2026-06-07-sibyl-phase-a-product-memory-core-v3`

## Scope

- System: Sibyl local memory.
- Mode: deterministic core memory test, no LLM extraction, no LLM answering.
- Companies: 50
- Questions: 120
- Questions per test: 20
- Tests: conversation_to_entity_memory, multi_session_agent_memory, provenance_auditability, conflict_and_correction, realistic_noise_and_scale, codex_agent_integration
- Stored records: 660
- Write calls: 660
- Model/API cost USD: 0.000000

## Summary

| Metric | Value |
| --- | ---: |
| Retrieval passed | 120 / 120 |
| Avg context rows | 5.70 |
| Avg context tokens | 697.08 |
| Ingest elapsed seconds | 1.081 |
| Retrieval elapsed seconds | 0.024 |
| DB size bytes | 876544 |

## Retrieval By Test

| Test | Retrieval passed |
| --- | ---: |
| conversation_to_entity_memory | 20 / 20 |
| multi_session_agent_memory | 20 / 20 |
| provenance_auditability | 20 / 20 |
| conflict_and_correction | 20 / 20 |
| realistic_noise_and_scale | 20 / 20 |
| codex_agent_integration | 20 / 20 |

## Interpretation

This Phase A1 run isolates the Sibyl memory engine. The dataset is deterministic and already structured as entities, state, journal events, and references. It verifies whether Sibyl can preserve and retrieve operational memory for companies, people, deals, provenance, corrections, noisy aliases, and agent tool continuity.

LLM extraction and answer-only runs should be evaluated separately so failures can be attributed cleanly.

## Failures

```json
[]
```