Memory-First Conversational Architecture for Chat Systems I’d like to start a discussion about an architectural direction for chat systems that could reduce reliance on extremely large models and very long context windows. The core idea is a memory-first conversational architecture. Instead of tr…

Many existing systems perform some kind of context “compaction”, e.g. Codex-CLI. I’ve dabbled with it in two ways with my project Discourse Chatbot: [image] GitHub - merefield/discourse-chatbot: An AI bot with RAG capability for Topics and Chat... An AI bot with RAG capability f…

You note Chat Systems but then have the topic tagged with API. [image] Should this topic be moved to the community category and retagged? I ask because there is a big difference between using the API and using ChatGPT. Depending on where you are headed this will be of value The key idea: durab…

Thanks for sharing this — your Discourse chatbot approach is very interesting. What I’m experimenting with is a bit different from a typical RAG pipeline. Instead of retrieving documents, I keep a structured memory layer in a database (SQLite) that stores conversation summaries, facts, and interac…

[image] EricGT: Thanks for pointing that out. You’re right — the topic is more about conversational system architecture than about a specific API implementation. I used the API tag mainly because the architecture I’m experimenting with is built around an API-based stack. The idea you mention…

Thanks for the suggestion. I originally posted this idea as feedback because I believe persistent conversational memory could significantly improve long-running chat systems. In my own experiments I implemented a prototype architecture where a central conversational model (~27B) works together wit…

Might I suggest Take a daily look at the commits on the OpenAI GitHub repository for Codex - You will notice that memory has been an active area for the last few weeks. Help the OpenAI Developer by giving him feedback - Memories in Codex · openai/codex · Discussion #12567 · GitHub Read the OpenAI …

And just to clarify at the start — I’m not throwing stones at OpenAI. I’ve worked with their models for a long time and respect what they built. My point is simply that they had a chance to be the first to move in this direction and remain the clear leader. I do follow the Codex repository and I se…

I would agree with much of your comments. Memory management via multicall and rag and langchain are the building blocks for a long horizon conversation and task (both). it allows the contextwindow to switch its attention as required. i suspect the reason that memory management long form conversation…

Memory-First Conversational Architecture as an Alternative to Long Context Windows

Prompting

merefield March 4, 2026, 11:13am 3

Additionally, I’d point you to this similar conversation here: Best practices for cost-efficient, high-quality context management in long AI chats

Topic		Replies	Views
A.U.R.O.R.A.: long-horizon continuity without steady context growth? Community project , projects , memory , ai-memory	9	369	June 11, 2026
The Elephant in the Room: Why No Persistent Conversational Memory in LLMs? Community feature-request , memory	68	5693	August 18, 2025
Persistent, Structured Memory Strategies Community chatgpt	14	1151	December 12, 2025
Best practices for cost-efficient, high-quality context management in long AI chats API context-window	4	4643	June 26, 2026
How do you maintain historical context in repeat API calls? API	28	95367	December 23, 2023

Memory-First Conversational Architecture as an Alternative to Long Context Windows

Related topics