FAMM: Solving the Memory Problem in LLM Agents

The Problem with LLM Memory

Large Language Models have a fundamental limitation: finite context windows. Even with models supporting 100K+ tokens, autonomous agents that run for hours or days inevitably face memory degradation:

• Context overflow — important early information gets pushed out
• Retrieval inefficiency — finding relevant past context becomes a needle-in-a-haystack problem
• Memory decay — the model's ability to reason about distant information degrades

Our Solution: FAMM

FAMM (Future-Aware Adaptive Memory Management) is a framework we designed to address these challenges. It was published on Zenodo with DOI 10.5281/zenodo.21168000.

Core Architecture

Incoming Context
    ↓
┌─────────────────────────────┐
│   Adaptive Memory Organizer │
│   (categorize + prioritize) │
└─────────────┬───────────────┘
              ↓
┌─────────────────────────────┐
│  Future-Aware Prioritizer   │
│  (predict future relevance) │
└─────────────┬───────────────┘
              ↓
┌─────────────────────────────┐
│  Intelligent Retriever      │
│  (multi-strategy retrieval) │
└─────────────┬───────────────┘
              ↓
┌─────────────────────────────┐
│  Dynamic Memory Optimizer   │
│  (compress + consolidate)   │
└─────────────────────────────┘

Key Innovation: Future-Aware Context Prioritization

Most memory management approaches for LLM agents are backward-looking — they store what happened and retrieve based on similarity to the current query. FAMM is different because it's forward-looking:

• It predicts which memories will be needed in the future based on the agent's current task trajectory
• It proactively elevates relevant memories before they're needed
• It compresses or discards information that's unlikely to be useful

Think of it like a chess player: you don't just remember past moves, you think about which past patterns are relevant to your next moves.

Memory Organization Tiers

FAMM organizes memory into adaptive tiers:

• 1.Active Memory — currently relevant context (always in the prompt)
• 2.Warm Memory — recently accessed, high future-relevance score
• 3.Cold Memory — compressed summaries of older interactions
• 4.Archive — highly compressed, retrievable only on specific queries

Intelligent Retrieval

Instead of single-strategy retrieval (e.g., just semantic similarity), FAMM uses multi-strategy retrieval:

• Semantic search — find contextually similar memories
• Temporal recency — prefer recent, relevant information
• Task dependency — retrieve memories that are prerequisites for current tasks
• Predictive loading — pre-fetch memories likely needed in the next 2-3 steps

Experimental Results

We evaluated FAMM against baseline approaches (sliding window, naive RAG, and fixed-buffer memory):

FAMM showed particular strength in:

• Multi-session tasks — maintaining context across conversation boundaries
• Complex reasoning chains — keeping track of intermediate results
• Long-term planning — remembering goals and constraints from early in the interaction

Why This Matters

As LLM agents move from single-turn chatbots to autonomous systems that operate for hours, days, or continuously, memory management becomes the bottleneck. FAMM is a step toward agents that can:

• Run customer support systems that remember user history across weeks
• Manage software projects with full context of every decision
• Conduct multi-day research with coherent, evolving understanding

Read the Full Paper

Download PDF

DOI: 10.5281/zenodo.21168000

By Suyash Vakhariya and Asmita Ipper