Stockages mémoire simples (clé-valeur)

La mémoire à long terme la moins coûteuse

Que stocker

Un petit stockage en mémoire

memory = {
    'user_name': 'Alice',
    'units': 'celsius',
    'language': 'tr'
}

def remember(key, value):
    memory[key] = value

def recall(key, default=None):
    return memory.get(key, default)

remember('city', 'Istanbul')
print("Recalled user_name:", recall('user_name'))
print("Recalled city:", recall('city'))
print("Recalled missing key:", recall('country', 'unknown'))

Injecter dans l'instruction système

def build_system_prompt(user_id):
    facts = load_facts(user_id)
    rendered = '\n'.join(f'- {k}: {v}' for k, v in facts.items())
    return f'''
You are AssistantBot.

Known facts about this user:
{rendered}

Use these to personalise replies.
'''

Utiliser Redis pour plusieurs utilisateurs

import redis
r = redis.Redis()

def remember(user_id, key, value):
    r.hset(f'mem:{user_id}', key, json.dumps(value))

def recall_all(user_id):
    raw = r.hgetall(f'mem:{user_id}')
    return {k.decode(): json.loads(v) for k, v in raw.items()}

Ou utiliser simplement Postgres

CREATE TABLE user_memory (
    user_id TEXT PRIMARY KEY,
    facts JSONB NOT NULL DEFAULT '{}'
);

-- Update one field:
UPDATE user_memory SET facts = facts || '{"units": "celsius"}' WHERE user_id = $1;

La mémoire comme outil

tools = [
    {'name': 'remember', 'description': 'Save a fact about the user', 'parameters': {...}},
    {'name': 'forget', 'description': 'Remove a stored fact', 'parameters': {...}}
]

# The agent can call remember('user_name', 'Alice') itself.
for t in tools:
    print(f"{t['name']}: {t['description']}")

Éviter le gonflement de la mémoire

Durées d'expiration et oubli progressif

r.hset(f'mem:{user_id}', 'last_topic', 'pricing')
r.expire(f'mem:{user_id}', 3600)  # 1 hour

Mémoire modifiable par l'utilisateur

Confidentialité : attention aux PII

De KV au vecteur

Quand utiliser la mémoire KV

Récapitulatif