प्रक्रियात्मक स्मृति (कौशल पुस्तकालय)

CREATE TABLE skills (
    id UUID PRIMARY KEY,
    name TEXT NOT NULL,
    description TEXT NOT NULL,
    embedding vector(1536),
    steps JSONB NOT NULL,           -- list of tool calls
    success_rate REAL DEFAULT 1.0,
    use_count INT DEFAULT 0,
    created_at TIMESTAMPTZ
);

extract_prompt = '''
The agent successfully completed: {user_request}

The trace was:
{trace}

Write this as a reusable skill with name, description, and step list.
Return JSON: {name, description, steps: [{tool, args_template}]}
'''
print(extract_prompt.strip())

skill = json.loads(llm.invoke(extract_prompt).content)
vec = embed(skill['description'])
conn.execute('''
    INSERT INTO skills (id, name, description, embedding, steps)
    VALUES (%s, %s, %s, %s, %s::jsonb)
''', (str(uuid.uuid4()), skill['name'], skill['description'], vec, json.dumps(skill['steps'])))

def find_skill(query):
    qvec = embed(query)
    rows = conn.execute('''
        SELECT name, steps FROM skills
        ORDER BY embedding <=> %s LIMIT 3
    ''', (qvec,))
    return rows

skill = {
    'name': 'file_expense_report',
    'steps': [
        {'tool': 'open_form', 'args': {'form_id': 'expense'}},
        {'tool': 'fill_field', 'args': {'field': 'amount', 'value': '{amount}'}},
        {'tool': 'fill_field', 'args': {'field': 'category', 'value': '{category}'}},
        {'tool': 'submit'}
    ]
}
import json
print(json.dumps(skill, indent=2))

candidates = find_skill(user_request)
prompt = f'''
You can use these existing skills if applicable:
{json.dumps(candidates, indent=2)}

Or you can plan a new sequence.
'''

UPDATE skills
SET use_count = use_count + 1,
    success_rate = (success_rate * use_count + (CASE WHEN %s THEN 1 ELSE 0 END)) / (use_count + 1)
WHERE id = %s;

DELETE FROM skills
WHERE (success_rate < 0.5 AND use_count >= 5)
   OR (use_count = 0 AND created_at < NOW() - INTERVAL '30 days');

low_level_skill = {'name': 'open_email', 'steps': [...]}
high_level_skill = {'name': 'process_inbox', 'steps': [
    {'skill': 'open_email'},
    {'skill': 'reply_or_archive'}
]}
print(low_level_skill)
print(high_level_skill)