aivery 0.1.0

AI memory that understands time — temporal-tree retrieval for LLM agents

Aivery Python SDK

AI memory that understands time — temporal-tree retrieval for LLM agents.

from aivery import Memory

m = Memory()
m.add("I love hiking on weekends", user_id="alice")
results = m.search("outdoor activities", user_id="alice")

Why Aivery?

Most memory systems store facts as a flat list. Aivery organizes memories into a temporal tree: related memories are parent/child, time flows from root to leaf, contradictions fork into branches. Retrieval walks the tree — you get a coherent branch of context, not a bag of facts.

Compared to mem0 on the LOCOMO benchmark (LLM judge score, higher = better):

System	LLM Score
mem0 (platform)	0.5571
Aivery (tree + heatmap, K=50)	0.6773
Aivery (wide retrieval, K=200→50)	0.8000

Installation

pip install aivery

Requires an Aivery API key. Point AIVERY_BASE_URL at https://api.aivery.systems or a self-hosted instance.

Quickstart

from aivery import Memory

m = Memory()

# Add from a plain string
m.add("I'm training for a marathon in April", user_id="alice")

# Add from a conversation
m.add(
    [
        {"role": "user", "content": "I just moved to San Francisco."},
        {"role": "assistant", "content": "Welcome! How are you finding it?"},
    ],
    user_id="alice",
)

# Search
results = m.search("where does alice live?", user_id="alice")
# [{"memory": "Alice lives in San Francisco", "score": 0.94, "id": "..."}]

# Get a context block for LLM injection
ctx = m.context("what do I know about alice?", user_id="alice")
# "- Alice is training for a marathon in April\n- Alice lives in San Francisco"

Drop-in OpenAI agent

import os
from openai import OpenAI
from aivery import Memory

openai = OpenAI()
memory = Memory()

def chat(message: str, user_id: str) -> str:
    ctx = memory.context(message, user_id=user_id, top_k=10)

    system = "You are a helpful assistant."
    if ctx:
        system += f"\n\nWhat you remember about the user:\n{ctx}"

    response = openai.chat.completions.create(
        model="gpt-4o-mini",
        messages=[
            {"role": "system", "content": system},
            {"role": "user", "content": message},
        ],
    )
    answer = response.choices[0].message.content

    memory.add(
        [{"role": "user", "content": message}, {"role": "assistant", "content": answer}],
        user_id=user_id,
    )
    return answer

Async

from aivery import AsyncMemory

async with AsyncMemory() as m:
    await m.add("I love jazz", user_id="alice")
    results = await m.search("music preferences", user_id="alice")

Configuration

Method	Description
Memory()	Reads AIVERY_BASE_URL, AIVERY_API_KEY, AIVERY_ORG_ID from environment
Memory(host="http://...")	Explicit host
Memory(api_key="aiv_...")	Explicit API key

Environment variables:

AIVERY_BASE_URL=https://api.aivery.systems   # default
AIVERY_API_KEY=aiv_...                       # required
AIVERY_ORG_ID=00000000-...                   # multi-tenant org scoping

Hosted API

Use the Aivery hosted API at https://api.aivery.systems — no setup required.

Reproducing the LOCOMO benchmark

The numbers in the table above were produced using the LOCOMO dataset and the evaluation harness from mem0. The benchmark/ directory in this repo contains the Aivery integration scripts — see benchmark/README.md for full instructions.

Cost breakdown: Ingestion and answer generation run through your Aivery plan (server-side LLM — no key needed on your end). The LLM judge step requires your own OpenAI key and costs ~$5 for the full 1,540-question set. A Pro tier or higher is recommended for ingestion throughput.

Self-hosting

Self-hosting docs and a one-command Docker setup are coming soon.

License

MIT