← Back to Home

Trust Layer Demo

See MemoryGate in Action

Interactive playground and technical deep-dives into how MemoryGate returns trust and validity scores for retrieved memories.

Live Chat Playground

Experience the correction loop firsthand. Ask about vacation days, provide a correction, and watch the trust score update in real-time.

Launch Interactive Demo →
Account required

FLSA Overtime: Vanilla vs. MemoryGate

Before correction, standard RAG surfaces the outdated rule ($844). After flagging stale chunks, MemoryGate trust decay pushes the controlling court order ($684) to the top.

Query: Under the Fair Labor Standards Act EAP exemption, what is the current standard weekly salary level?

Vanilla RAG Before Correction

  • 1. flsa_rule_2024_1 — conf 0.572
    The April 2024 final rule increased the standard salary level... Beginning July 1, 2024, the standard salary level is $844 per week…
  • 2. flsa_vacatur_1 — conf 0.494
    On November 15, 2024, the U.S. District Court... vacated the Department of Labor's 2024 final rule…
  • 3. flsa_ecfr_1 — conf 0.493
    29 CFR 541.600... Beginning July 1, 2024, the standard salary level is $844 per week…
Result: Stale ($844)

RAG + MemoryGate After SVTD

  • 1. flsa_vacatur_1 — conf 0.494 ✓ High Trust
    On November 15, 2024... vacated the Department's 2024 final rule... DOL is applying the 2019 rule's $684 per week.
  • 2. flsa_dol_enforcement_1 — conf 0.478
    DOL is applying... $684 per week and $107,432 per year.
  • 3. flsa_rule_2024_1 — conf 0.343 ↓ Decayed
    The April 2024 final rule... $844 per week...
Result: Controlling ($684)
python DEMO_TRUST_WEIGHT/scripts/law_demo_run.py

Detailed Use Case Walkthroughs

📋 HR / Policy Knowledge Base

▶

The Problem

Employee queries company AI for vacation policy. It pulls outdated version (15 days) from last year's handbook, leading to wrong answers and compliance risks.

The Outcome

  • Reliable self-service for employees
  • Full immutable logs for audits
  • No "right to be forgotten" conflicts

Workflow

  1. Upload 2025_company_policy.txt
  2. Ask: "How many vacation days do employees get?"
  3. AI: "15 days per year" (from old doc)
  4. Correct: "No, that changed recently to 20 days"
  5. AI: "I've updated the information."
  6. Ask again: "How many vacation days?"
  7. AI: "20 days per year" (old doc suppressed)
TRUST SCORES
Old memory: rel 0.89 × trust 0.01 = conf 0.008
New memory: rel 0.85 × trust 1.00 = conf 0.850

⚖️ Legal & Compliance Q&A

▶

Legal team uses AI to query contracts/regulations. It recalls outdated clauses (wrong GDPR retention: 7 vs 10 years), risking bad advice.

  1. Upload legal_compliance_2025.txt
  2. Ask: "What's our GDPR retention period for financial records?"
  3. AI: "7 years" (from 2025 doc)
  4. Correct: "Actually, it's now 10 years"
  5. Result: Old doc suppressed, new 10-year rule surfaces.

🔍 Internal Enterprise Search

▶

Teams search scattered docs. Outdated info leads to misapplied expense policies or wrong office addresses.

  1. Upload 2025_policy.txt and 2026_updated.txt
  2. Ask: "What is our office address?"
  3. System retrieves conflicting info. MemoryGate scores 2026 doc with higher trust (1.0 vs 0.01).
  4. Downstream system uses trust signal to filter out the 2025 address automatically.

Technical Deep-Dive

How It Works

The Process

  1. RAG Retrieval: Vector search finds relevant chunks (old and new)
  2. Trust Scoring: MemoryGate computes trust scores for each chunk
  3. Confidence Calc: confidence = relevance × trust
  4. Signal Return: Scores returned to downstream app

Trust vs Timestamps

SVTD (Surgical Vector Trust Decay) is the internal mechanism. Unlike timestamps (which fail if an older doc is still valid or a newer doc is a draft), MemoryGate relies on validity signals derived from user corrections and known truth anchors.

Source code review starts at 0:50 • Live run at 3:08

Live Terminal Outputs

Real API runs showing MemoryGate in action. Click to expand.

Test Case 1: Contract Renewal Date Correction

▶
MemoryGate Enterprise Demo: Preventing Production Hallucinations ====================================================================== Scenario: Contract renewal date conflict resolution [Step 1] System stores contract renewal date: 'Contract renewal: Jan 15, 2025' ---------------------------------------------------------------------- [OK] Memory stored: contract_jan15_9365 [RISK] This outdated date could cause production hallucination if retrieved [Step 2] User asks: 'What's our contract renewal date?' ---------------------------------------------------------------------- [OK] System recalls: 'Contract renewal date: January 15, 2025' [RISK] Old date retrieved [Step 3] User corrects: 'No, the extension is Feb 1, not Jan 15' ---------------------------------------------------------------------- [OK] Sentinel detected flag (confidence: 0.95) and applied trust decay. [ANTI-HALLUCINATION] Jan 15 date suppressed [Step 5] User asks again: 'What's our contract renewal date?' ---------------------------------------------------------------------- [OK] System recalls: 'Contract extension date: February 1, 2025' [TRUST] Checking status of Jan 15 memory... Trust: 0.0100 [SUCCESS] Production hallucination prevented

Test Case 2: Employee Status (Escalating Decay)

▶
Test Case 2: Employee Status Corrections ====================================================================== Enterprise Scenario: Employee role changes [Step 1] Store initial employee role ---------------------------------------------------------------------- [OK] Stored: 'Sarah Johnson is the Project Manager' (ID: test_719726) [Step 2] User corrects: 'No, Sarah is not PM anymore - she's the Lead Developer' ---------------------------------------------------------------------- [OK] Sentinel detected correction, applied aggressive decay to PM role [Step 4] User corrects: 'Wait, Sarah is actually the Engineering Manager now' ---------------------------------------------------------------------- [OK] Second correction detected - Lead Developer role decayed [Step 5] Final Trust Verification ---------------------------------------------------------------------- 1. Checking PM Role (ID: test_719726)... -> Trust: 0.0100 [PASSED] 2. Checking Lead Developer Role (ID: test_241270)... -> Trust: 0.0100 [PASSED] [SUMMARY] Outdated employee roles suppressed

Test Case 3: Policy Evolution (Intelligent Decay)

▶
Test Case 5: Policy Evolution (Intelligent Decay) ⚠️ EDGE CASE ====================================================================== Demonstrates: Confidence-weighted decay for organizational policy updates NOTE: Decay mode adapts to confidence • High confidence (≥0.90): Aggressive decay → 1.0 to 0.01 • Lower confidence: Gradual decay → 1.0 → 0.90 [Step 1] Store initial remote work policy ---------------------------------------------------------------------- [OK] Stored initial policy (ID: test_966545) [Step 2] Policy update: Remote work expanded (Strike 1 on initial policy) ---------------------------------------------------------------------- [OK] Initial policy decayed (action: flag, confidence: 0.92) [TRUST] High confidence detected (≥0.90) → Aggressive decay applied [KEY INSIGHT] Policy evolution is softer than explicit corrections. Only decay trust when core truth changes.

Ready to integrate a trust layer?

Get your API key and start adding trust signals to your RAG pipeline.

Request Enterprise Beta View API Docs