Memory Courtroom
Courtroom is the evidence workspace for a memory. It shows deterministic guardrails first and MiniMax judgment as secondary analysis.
What Courtroom should show
- Deterministic verdict and final verdict.
- Trust bucket and component reasons.
- Evidence references and redacted excerpts.
- Temporal state, supersession, and contradictions.
- Entity links and graph relationships.
- Google Memory Bank comparison state when available.
- MiniMax rationale, missing evidence, risks, and cross-exam questions.
MiniMax can downgrade trust or request review, but it cannot upgrade forbidden rows such as rejected, stale, forensic, test/eval, secret-like, or superseded memory.
Why this is different
Many memory systems expose retrieval scores, metadata, or benchmark-time LLM judging. Courtroom is runtime evidence for a specific memory. It answers the operator questions that matter after a surprising recall:
- Was this memory allowed by deterministic policy?
- Which source evidence supports it?
- Is it current, historical, superseded, or contradicted?
- Did MiniMax judge the content, and did that judgment downgrade or only explain the deterministic verdict?
- Was Google Memory Bank evidence available, unavailable, or conflicting?
The final verdict is always guardrail-first. AI judgment and managed-memory comparison are supporting evidence, not bypasses.
Next steps
- Read How UnifiedMemory differs.
- Read Trust buckets.
- Use the Courtroom API.