alt Hacker NewsI scrolled straight to section 2.1 "Definition and Characteristics of LLM agents" to find out which of the many definitions of "agent" they are using here.
They went for LLM + short-term and long-term memory + planning + tool using + action execution.
Presumably "planning" here is covered by any LLM that can do "think step by step" reasonably well?
It wasn't clear to me what the difference between "tool using" and "action execution" was.
I haven't seen a definition that specifically encompasses both short- and long-term memory before. They say:
> This dual memory system allows agents to maintain conversation continuity while building knowledge over time.
So presumably, this is the standard LLM chat conversation log plus a tool that can decide to stash extra information in a permanent store - similar to how ChatGPT's memory feature worked up until about four weeks ago.
Replies
Accurate memory access across memory types is still not really solved. That is the issue. Most agent frameworks from the main model providers are still quite spotty.
Image of a table outlining where the major frameworks are: https://substackcdn.com/image/fetch/w_1272,c_limit,f_webp,q_...
Here is also an article I wrote last year on the different types of memory: https://open.substack.com/pub/jdsemrau/p/memory-and-knowledg...
> It wasn't clear to me what the difference between "tool using" and "action execution" was.
doing a lot of inference here, but could be a separation between -read- tool kinda actions, and -write/execute- (like running code/sending an email, etc)
a bit weird from a coding perspective but idk
> similar to how ChatGPT's memory feature worked up until about four weeks ago
What happened four weeks ago?
> Presumably "planning" here is covered by any LLM that can do "think step by step" reasonably well?
mild disagree. 1) externalizing the plan and letting the user audit/edit the plan while its working is "tool use", yes, but a very specialcase kind of tool use that, for example, operator and deep research use Temporal for. ofc we also saw this with Devin/Manus and i kinda think they're better 2) there is a form of primitive tree search that people are doing where they can spam out several different paths and run it a few steps ahead to gain information about optimal planning. You will see this with morph's launch at AIE. 3) plan meta reflection and reuse - again a form of tool use, but the devin and allhands folks have worked on this a lot more than most.
my criticism of many agent definitions is that they generally do not take memory, planning, and auth seriously enough, and i think those 3 areas are my current bets for "alpha" in 2025.
> I haven't seen a definition that specifically encompasses both short- and long-term memory before.
here
- https://docs.mem0.ai/core-concepts/memory-types#short-term-m...
- https://x.com/swyx/status/1915128966203236571