compass-event-memory

name	compass-event-memory
title	Memory Matters More: Event-Centric Memory as a Logic Map for Agent Searching and Reasoning
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2601.04726
keywords	["agent-memory","event-graph","structured-retrieval","reasoning-over-memory","long-horizon-planning"]
description	Organize agent memory as an event graph with explicit logical relationships rather than flat embeddings. Framework incrementally segments experiences into events and links them through causal, temporal, and logical relations. Enables agents to navigate memory as a logic map for goal-directed searching and structured reasoning, improving performance on multi-hop reasoning and long-horizon planning tasks.

Problem

Current LLM agent memory systems suffer from three limitations:

1. Shallow Retrieval: Similarity-based lookup returns semantically related but logically disconnected memories
2. Lost Context: Flat memory stores lose the narrative flow and causal relationships between experiences
3. Poor Long-Horizon Planning: Agents struggle to navigate sequences of memory spanning multiple steps because relationships aren't captured
4. Inefficient Search: Finding the "right" memory for multi-step reasoning requires expensive retrieval iterations

Agents need memory that captures not just what happened, but how events connect logically.

Solution

CompassMem (Memory Matters More) organizes memory as an Event Graph:

1. Event Segmentation: Incrementally break experience sequences into discrete, self-contained events
2. Logical Relations: Link events through explicit relationships:
   • Causal: Event A led to Event B
   • Temporal: Event A happened before Event B
   • Conditional: Event B only occurred because Event A established preconditions
   • Goal-Oriented: Events form chains progressing toward specific goals
3. Graph Navigation: Agents traverse this logic map to gather relevant experiences, moving from high-level goals to supporting details

When to Use

• Multi-hop Reasoning: Tasks requiring agents to chain knowledge across multiple memory segments
• Long-Horizon Planning: Agents coordinating actions over extended sequences
• Dialogue Agents: Maintaining context through multi-turn conversations with rich interaction history
• Collaborative Agents: Multiple agents sharing structured memory about past interactions
• Question Answering: Retrieving not just facts but the causal chain leading to conclusions

When NOT to Use

• For simple retrieval tasks (single-fact lookup is more efficient with embeddings)
• In real-time systems where event segmentation overhead is prohibitive
• When memory access patterns are random (graph structure provides no advantage)
• For agents with limited interaction history (insufficient events to build meaningful graphs)

Core Concepts