Agentic Learning Without Retention

zenodo.org · 2026-10-05 · by Long Nguyen ·generic
We keep trying to build generalist agents by giving them more stuff (more memory, more skills). Hoping all of that somehow turns them into better decision makers! The problem with most agents is that they don't actually need more memory. The current structure is broken; save the conversation > save the decisions > retrieve from vector db > and hope the right thing comes back. That might help an agent continue the same task, but it doesn't really help the next agent. That system design worked for low to mid level tasks, but it will not work for high level decision making. It becomes a problem as agents get more capable. A long history is not the same thing as understanding. Most of what an agent experiences is probably specific to that situation, mixed with wrong assumptions, temporary conditions, and noise. Keeping all of it means the agent has to keep carrying around an ever-growing pile of stuff and somehow figure out what still matters. This paper takes almost the opposite approach. Instead of trying to retain experience, it tries to extract whatever survives from experience and throw the rest away. The agent doesn't need to remember every time it solved a problem. It needs to figure out what those experiences reveal about how things actually work, try to destroy those explanations, and only keep the parts that continue to survive. That is the idea behind  A true generalist shouldn't need to remember every experience it had, or even know much about the domain it is working in. It shouldn't need to recall the actually domain knowlege, that can be outsourced cheaply. What the agent actually needs to carry from one experience to another is an understanding of why things worked or failed. This paper proposes a new type of architecture: the active agent has many experiences, some in totally different domains. a slow reflective layer called "Shaman" looks at those experiences when the agent quota hasn't finished. it makes a hypothses about the underlying mechanics of those experiences. then it tries to kill those hypothses. The hypothses that survive long enough get used in forming general principles.The principles that are backed by more than 1 hypothses get more attention.The star principles are used to build a worldmodel that shaped new agents. The individual experiences are deleted completely.  Experience → Hypothesis → Refutation → Principle → World Model.    

Agentic Learning Without Retention

Published August 3, 2026 | Version 26.9.12

Preprint Open

Agentic Learning Without Retention

Authors/Creators

  • Shoman, Salah

    • Description

    We keep trying to build generalist agents by giving them more stuff (more memory, more skills). Hoping all of that somehow turns them into better decision makers!

    The problem with most agents is that they don't actually need more memory. The current structure is broken; save the conversation > save the decisions > retrieve from vector db > and hope the right thing comes back. That might help an agent continue the same task, but it doesn't really help the next agent. That system design worked for low to mid level tasks, but it will not work for high level decision making.

    It becomes a problem as agents get more capable. A long history is not the same thing as understanding. Most of what an agent experiences is probably specific to that situation, mixed with wrong assumptions, temporary conditions, and noise. Keeping all of it means the agent has to keep carrying around an ever-growing pile of stuff and somehow figure out what still matters.

    This paper takes almost the opposite approach. Instead of trying to retain experience, it tries to extract whatever survives from experience and throw the rest away. The agent doesn't need to remember every time it solved a problem. It needs to figure out what those experiences reveal about how things actually work, try to destroy those explanations, and only keep the parts that continue to survive. That is the idea behind

    A true generalist shouldn't need to remember every experience it had, or even know much about the domain it is working in. It shouldn't need to recall the actually domain knowlege, that can be outsourced cheaply. What the agent actually needs to carry from one experience to another is an understanding of why things worked or failed.

    This paper proposes a new type of architecture:

  • the active agent has many experiences, some in totally different domains.
  • a slow reflective layer called "Shaman" looks at those experiences when the agent quota hasn't finished.
  • it makes a hypothses about the underlying mechanics of those experiences.
  • then it tries to kill those hypothses.

The hypothses that survive long enough get used in forming general principles. The principles that are backed by more than 1 hypothses get more attention. The star principles are used to build a worldmodel that shaped new agents.

The individual experiences are deleted completely.

Experience → Hypothesis → Refutation → Principle → World Model.

Files

###

LWR 26.9.pdf

###

Files (70.4 kB)

| Name | Size | Download all | | ------------------------------------------------------------------------------------------------------------------------ | ------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | LWR 26.9.pdf md5:295751a18b94a4dc13a5e1493e9bb7cd | 70.4 kB | Preview Download |

Additional details

2026-08-03

2026-10-03

<https://github.com/sshoman/lwr>

Jump up

zenodo.org · 6 images · 42 links

Comments (0)

No comments yet