Pinecone vs Kronvex
Base vectorielle vs API Mémoire

PINECONE — build it yourselfimport openai
from pinecone import Pinecone
from datetime import datetime
import math

pc = Pinecone(api_key="pc-...")
index = pc.Index("agent-memories")
oai = openai.OpenAI()

def get_embedding(text):
    r = oai.embeddings.create(
        model="text-embedding-3-small",
        input=text
    )
    return r.data[0].embedding

def remember(agent_id, text):
    vec = get_embedding(text)
    index.upsert(vectors=[{
        "id": f"{agent_id}_{hash(text)}",
        "values": vec,
        "metadata": {
            "text": text,
            "agent_id": agent_id,
            "ts": datetime.utcnow().isoformat(),
            "access_count": 0
        }
    }], namespace=agent_id)

def inject_context(agent_id, query, n=5):
    vec = get_embedding(query)
    results = index.query(
        vector=vec, top_k=20,
        namespace=agent_id,
        include_metadata=True
    )
    # Manual confidence scoring: you build this
    now = datetime.utcnow()
    scored = []
    for m in results.matches:
        sim = m.score
        ts = datetime.fromisoformat(m.metadata["ts"])
        age_days = (now - ts).days
        recency = 1 / (1 + math.exp(-(age_days - 30) / 5))
        freq = math.log1p(m.metadata.get("access_count", 0))
        score = sim*0.6 + recency*0.2 + freq*0.2
        scored.append((score, m.metadata["text"]))
    scored.sort(reverse=True)
    top = [t for _, t in scored[:n]]
    # Format as prompt block — also your job
    return "Memory context:\n" + "\n".join(
        f"- {t}" for t in top
    )

# ~50 lines of boilerplate, just for basic memory