# LLM Reasoning and Planning with Memory, Tools and Action · Yila

> A hub-and-spoke framework with an LLM reasoning and planning block at its centre: multimodal inputs on the left, two memory stores above, four reasoning strategies on the right, and separate tool and action branches below it.

URL: https://yila.ai/templates/llm-agent-network-figure

A dark blue central box reads LLM Reasoning and Planning beneath a brain-circuit icon. On the left, a panel of text, voice and picture inputs feeds it through a blue arrow. Above, short-term memory in blue and long-term memory in red each connect to it with a two-way arrow. On the right, a red arrow leads to chain of thought, task decomposition, plan selection and reflection. Below, the flow splits into a blue Tools box serving computing, web and calendar, and a red Action box leading to coding. Readers should see that the model plans, memory supplies context, and tools and actions carry the plan into the outside world.

- **Central LLM block**: The hub of the whole figure; relabel it with the model or reasoning method your paper actually uses.
- **Multimodal inputs**: Text, voice and picture in the left panel; add or remove input types to match what your system accepts.
- **Input arrow**: The blue arrow from the input panel to the hub; note any encoder or preprocessing step that sits on this path.
- **Short- and long-term memory**: Blue and red database icons above the hub; say which corresponds to the context window and which to a vector store.
- **Reasoning strategies**: Chain of thought, task decomposition, plan selection and reflection, each with an icon; keep only those your paper uses.
- **Tools branch**: Computing, web and calendar icons hang under the blue box; swap in the actual tools your agent can call.
- **Tools-to-action arrow**: A dashed arrow runs from Tools to Action; explain how a tool result triggers an action in your design.
- **Action branch**: The red Action arrow ends in empty space; label the environment, file system or device it acts on.

- Paper mechanism figure, Defense slides, Course handout, Lab website

- **Missing observation loop**: Information flows only from inputs to action, with no path returning results to the model. Reviewers will note that an agent rests on an act-observe cycle and will ask for a feedback arrow to close it.
- **Dangling action arrow**: The red arrow from Action points at nothing, and the coding icon also hangs from the tools line. Clarify where action ends and tool use begins, and draw what the action actually affects. One arrow fixes it.
- **Memory read and write**: Two-way arrows alone do not explain memory. Expect questions about when items are written to long-term memory, how they are retrieved, and how limited capacity and forgetting are handled. Label both arrows.
- **Where strategies sit**: The four strategies are drawn as outputs of the model, though they organise its reasoning rather than follow it. Consider showing them inside the central block instead of downstream of it.

## FAQ

### Can this LLM agent framework go straight into my paper?

Use it as a draft for the methods section. It is AI-generated, so the modules and arrows illustrate a common pattern rather than your implementation. Add the feedback loop, fix the dangling arrow and rename modules after your real components first, then have a co-author compare it with the code.

### Can I swap modules or icons?

Icons cannot be selected and replaced one at a time, and the PNG is not a vector graphic. Describe the change in a Yila chat, such as splitting memory into episodic and semantic stores or adding a second agent, and the figure is regenerated with the same colour scheme.

### What will the file be?

An AI-generated PNG in landscape, with inputs, memory, strategies and outputs arranged around the central block. For vector artwork or text that matches your paper's font, redraw the structure in illustration software using this layout as the plan.

### Does it cost money?

Viewing the page and sample costs nothing. Generating and revising figures in Yila consumes credits; new users receive 2,500 credits when they sign up, which covers a few rounds of edits to the layout and the module names. Opening sample pages never uses credits.

### Can the framework become a retrieval-augmented setup?

Yes. Describe the retriever, knowledge base and reranker and the order in which they connect, and Yila replaces the memory section with a retrieval pipeline. Say whether the tool branch should stay, since many retrieval-augmented systems do not call external tools.

### Can I compare two agent frameworks in one figure?

You can ask for two frameworks side by side with identical icons and colours for shared modules and a highlight around what differs. If the difference is small, a single framework with a dashed box around the changed part usually reads better. Keep labels identical across both sides.

---

- [LLM Reasoning and Planning with Memory, Tools and Action](https://yila.ai/templates/llm-agent-network-figure)
- [llms.txt](https://yila.ai/llms.txt)
