Long-term evaluation of agents

An agent with a memory should become better for you over time: remember your preferences, not repeat mistakes, build on earlier work. An eval per answer does not see that. You need session series.

The setup:

• Simulated users with a hidden profile (goals, preferences, knowledge) who interact over 10–50 sessions. Measure whether the agent has learnt the profile: fewer questions, the right adaptation, no repeated mistakes.
• Memory probes: after session k, ask about something from session j < k. Measure the recall and whether the agent uses it unprompted.
• Measures over time: a curve of task success per session — the slope is the interesting part.
• Harm control: false memories, mix-ups between users, a memory that makes the agent worse (locking a misconception in).
• A control: the same agent without memory. The difference is the memory's value.

In AI-grafen the natural metric is «mastery per hour» over weeks, not points per exercise.

import numpy as np

def run_series(agent, profile, n_sessions, tasks, rng):
    """Returns the success per session and the memory probe hits."""
    success, probes = [], []
    for k in range(n_sessions):
        t = tasks[k % len(tasks)]
        answer = agent.session(t, simulated_user(profile, rng))
        success.append(t.judge(answer))
        if k > 0:
            fact = profile.fact_revealed_in(session=rng.integers(0, k))
            probes.append(fact.lower() in agent.ask(f"What do you know about {fact.subject}?").lower())
    return np.array(success), np.array(probes)

def slope(y):
    x = np.arange(len(y)); return float(np.polyfit(x, y, 1)[0])

# the report: the slope with memory vs without, the probe recall, the share of false memories
# (a probe about facts that were NEVER said)