autourgos-summary-memory

LLM-compressed rolling summary memory for Autourgos agents. Keeps the last N messages in full. When the buffer overflows, older messages are fed to an LLM for compression and merged into a rolling summary. The summary + recent messages are both included in every LLM prompt.

from autourgos_summary_memory import SummaryBufferedMemory
from autourgos_openaichat import OpenAIChatModel
from autourgos_agent import Agent

summarizer_llm = OpenAIChatModel(model="gpt-4o-mini")  # needs OPENAI_API_KEY set
memory = SummaryBufferedMemory(llm=summarizer_llm, max_messages=10)
agent = Agent(llm=summarizer_llm, memory=memory)
agent.invoke("Start a long research task...")

Features

• Rolling LLM compression — older messages are summarized, not dropped, so context survives long conversations
• Works without an LLM too — falls back to verbatim concatenation, still prevents unbounded growth
• Pairs with autourgos-summarizer — that package compresses the agent's reasoning scratchpad, this one compresses conversation history; different jobs, safe to use together

Table of Contents

• Install
• Quick Start
• Without an LLM
• Parameters
• What format_for_llm Returns
• License

Install

pip install autourgos-summary-memory

Quick Start

from autourgos_summary_memory import SummaryBufferedMemory
from autourgos_openaichat import OpenAIChatModel
from autourgos_agent import Agent

# Use a cheap model for summarization
summarizer_llm = OpenAIChatModel(model="gpt-4o-mini")  # needs OPENAI_API_KEY set
my_llm = summarizer_llm  # or any other chat-model instance for the main agent

memory = SummaryBufferedMemory(
    llm=summarizer_llm,
    max_messages=10,   # keep last 10 messages in full; compress the rest
)
agent = Agent(llm=my_llm, memory=memory)
agent.invoke("Start a long research task...")

Without an LLM

If no LLM is provided, overflow messages are concatenated verbatim (no AI compression). Still prevents unbounded growth:

memory = SummaryBufferedMemory(max_messages=10)

Parameters

What format_for_llm Returns

--- Summary of Past Conversation ---
[compressed history here]
------------------------------------

--- Recent Conversation Context ---
[2024-...] user: latest messages
[2024-...] agent: in full
-----------------------------------

License

Apache License 2.0, Copyright (c) 2026 Jitin Kumar Sengar