Memory function description added to paper

paper/conference_101719.tex

@@ -161,7 +161,7 @@ Explicit resource management is central, moving beyond simple API cost tracking.
     \item \textbf{Memory Usage:} Footprint of active Employee agents (\% of allocated budget).
     \item \textbf{Agent Concurrency:} Count of concurrently active agents.
 \end{itemize}
-Metrics are monitored against predefined \textbf{budget limits}. Actions (like hiring) exceeding limits (e.g., >90\% memory, exceeding max concurrency) are prevented. This ensures operation within constraints, crucial for limited resources or strict budgets.
+Metrics are monitored against predefined \textbf{budget limits}. Actions (like hiring) exceeding limits (e.g., $>$90\% memory, exceeding max concurrency) are prevented. This ensures operation within constraints, crucial for limited resources or strict budgets.
 
 \subsection{Tool Utilization and Autonomous Creation}
 \label{subsec:tooling}
@@ -176,6 +176,13 @@ A distinctive feature is \textbf{integrated, autonomous tool creation}. If the C
 \end{enumerate}
 This allows HASHIRU to dynamically extend its functional repertoire, tailoring capabilities to tasks without manual intervention, enabling greater autonomy and adaptation.
 
+\subsection{Memory Function: Learning from Experience}
+\label{subsec:memory}
+
+To enable HASHIRU agents to learn from past interactions and rectify previous errors, a \textbf{Memory Function} is incorporated. This function stores records of significant past events, particularly those involving failed attempts or suboptimal outcomes, acting as a historical log of experiences. When the system encounters a new problem or a recurring challenge, it queries this memory store to retrieve relevant past situations and their outcomes.
+
+Memory retrieval is based on semantic similarity between the current context (e.g., task description, recent actions, error messages) and the stored memory entries. We utilize embeddings generated by the \textbf{all-MiniLM-L6-v2} model \cite{wang2020minilmdeepselfattentiondistillation} to represent both the query and the stored memories in a high-dimensional vector space. Relevance is determined by calculating the \textbf{cosine similarity} between the query embedding and each memory embedding. Memories exceeding a predefined similarity threshold are retrieved and provided to the CEO agent (or relevant Employee agents) as contextual information. This allows the system to draw upon past experiences, understand why previous approaches failed, and potentially adjust its strategy to avoid repeating mistakes, thereby improving performance and efficiency over time.
+
 \section{Experimental Setup}
 \label{sec:experiments}
 

paper/references.bib

@@ -19,6 +19,16 @@
   year={2024}
 }
 
+@misc{wang2020minilmdeepselfattentiondistillation,
+      title={MiniLM: Deep Self-Attention Distillation for Task-Agnostic Compression of Pre-Trained Transformers}, 
+      author={Wenhui Wang and Furu Wei and Li Dong and Hangbo Bao and Nan Yang and Ming Zhou},
+      year={2020},
+      eprint={2002.10957},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL},
+      url={https://arxiv.org/abs/2002.10957}, 
+}
+
 @misc{gemini25flash,
   title = {Gemini 2.5 Flash: Model Card, API, and Announcement},
   author = {{Google DeepMind} and {Google AI}},