# How much of today's agent memory stack is a workaround for the model architecture? — retained record

> Retained reddit record from reddit:artificial.

Canonical: https://highsignal.app/data/records/2ee0d93a9167742e

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:artificial)  published 2026-10-14 14:34 ingested 2026-10-15 02:37

[original source ↗](https://www.reddit.com/r/artificial/comments/1x5sul4/how_much_of_todays_agent_memory_stack_is_a/)

## retained text



```
Much of today's agent stack is designed around the fact that model calls don't automatically carry forward everything learned during previous interactions. We retrieve conversation history, maintain external memory stores, summarize old context, compress prompts, orchestrate tools and decide what to inject again at each step.

Reading about recurrent-memory architectures got me thinking: what if some of that responsibility moved into the model itself? Instead of reconstructing the relevant context for every call, an agent could carry forward useful task-specific associations as an internal state.

Architectures built around recurrent state from day one, such as BDH-CQ, make that possibility interesting. In BDH-CQ, examples update an internal task memory without changing the model's weights, while a separate latent workspace handles reasoning without generating intermediate thought tokens. If that approach extends to long-running agents, the bigger shift might be less "cheaper reasoning" and more "less infrastructure devoted to reminding the model what it already learned." Retrieval would still matter for current documents, authoritative records, and information the model hasn't seen. But retrieving new information and reconstructing previously processed context are different jobs. For people building agents: which parts of your memory stack would you trust the model to handle internally, and which would you keep external regardless?
```



document fetched 2026-10-15 00:20 · canonical https://reddit.com/r/artificial/comments/1x5sul4/how_much_of_todays_agent_memory_stack_is_a
