LLM memory decay benchmark: add EdTech domain scenario (student/teacher conversation memory)

[Neal006]

Domain

Education / EdTech. A student AI tutor that needs to remember a learner's profile across a long tutoring session:

• Subject performance (struggling with calculus derivatives, strong at statistics)
• Learning style (prefers visual examples, needs step-by-step breakdowns)
• Session goals (preparing for exam on March 15)
• Past mistakes (confused about chain rule at T=20, corrected at T=35)

Why This Matters for LLM Memory Evaluation

The BENCHMARK_FACTS in simulator/facts.py only covers a personal assistant scenario (name, city, occupation). An EdTech scenario tests:

• Hierarchical facts: subject → topic → subtopic
• Evolving understanding: student's mastery level changes over the session
• Multi-update facts: the same concept can be "not understood" → "partial" → "mastered"

Implementation

# simulator/edtech_facts.py
EDTECH_FACTS = [
    Fact("student_name",    "Priya Nair",              injected_at=0),
    Fact("subject",         "calculus",                injected_at=1),
    Fact("weak_topic",      "chain rule",              injected_at=2,
         updated_at=35, updated_value="integration by parts"),
    Fact("exam_date",       "March 15",                injected_at=3),
    Fact("learning_style",  "visual learner",          injected_at=4),
    Fact("grade_target",    "A",                       injected_at=5),
    Fact("last_score",      "72%",                     injected_at=7,
         updated_at=60, updated_value="84%"),
    Fact("preferred_pace",  "slow with examples",      injected_at=9),
]

Acceptance Criteria

• simulator/edtech_facts.py with EDTECH_FACTS list
• python main.py --scenario edtech flag (or just use --facts edtech)
• Update README.md results table with EdTech scenario numbers
• CI passes with new scenario