コンテキストウィンドウにおける短期記憶

messages = [
  {'role': 'system', 'content': 'You are helpful.'},
  {'role': 'user',   'content': 'My name is Alice.'},
  {'role': 'assistant', 'content': 'Nice to meet you, Alice.'},
  {'role': 'user',   'content': 'What is my name?'}
]
# Model replies: 'Alice'
for m in messages:
    print(f"{m['role']}: {m['content']}")
print("Model replies: Alice")

MAX_TURNS = 20  # 10 user + 10 assistant

def trim(messages):
    system = messages[0]
    rest = messages[1:]
    return [system] + rest[-MAX_TURNS:]

demo_messages = [{'role': 'system', 'content': 'sys'}] + [
    {'role': 'user' if i % 2 == 0 else 'assistant', 'content': f'msg {i}'} for i in range(30)
]
trimmed = trim(demo_messages)
print(f"Original messages: {len(demo_messages)}")
print(f"Trimmed messages: {len(trimmed)}")

import tiktoken
enc = tiktoken.encoding_for_model('gpt-4o-mini')

def trim_by_tokens(messages, max_tokens=8000):
    out = [messages[0]]   # keep system
    tokens_left = max_tokens - len(enc.encode(messages[0]['content']))
    # walk backwards from newest
    for m in reversed(messages[1:]):
        cost = len(enc.encode(m['content'])) + 4
        if cost > tokens_left:
            break
        out.insert(1, m)
        tokens_left -= cost
    return out

# Bad: cuts after assistant call but before tool result
# Good: trim whole turns (user + all responses)
print("Bad: cuts after assistant call but before tool result")
print("Good: trim whole turns (user + all responses)")

SYSTEM_PROMPT = 'You are a helpful assistant.'
messages = [
    {'role': 'system', 'content': SYSTEM_PROMPT},
    {'role': 'system', 'content': 'USER PREFERENCE: Celsius units'},
    {'role': 'user', 'content': 'What is the weather in Paris?'},
    {'role': 'assistant', 'content': 'It is 21C and sunny.'},
]
print('Pinned system messages:')
for m in messages:
    if m['role'] == 'system':
        print(' -', m['content'])
print('Total messages:', len(messages))