💬 NLP

GRAVITY: Architecture-Agnostic Structured Anchoring for Long-Horizon Conversational Memory

The paper introduces GRAVITY, an architecture-agnostic plug-and-play module that enhances long-horizon conversational memory by extracting and injecting structured relational, temporal, and thematic representations into prompts, thereby significantly improving reasoning accuracy across diverse language models without requiring architectural modifications.

Original authors: Yushi Sun, Bowen Cao, Dong Fang, Lingfeng Su, Wai Lam

Published 2026-05-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yushi Sun, Bowen Cao, Dong Fang, Lingfeng Su, Wai Lam

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are talking to a very smart friend who has read thousands of books and had millions of conversations. This friend is great at answering questions, but they have a problem: their memory is like a giant, messy attic.

When you ask, "What did we talk about last month regarding my dog?" they can find the right box in the attic. But when they pull it out, it's just a pile of loose papers, receipts, and scribbled notes. They have to read through the whole mess to figure out who the dog is, when you met the vet, and what the vet said. They have to do all that heavy lifting while they are trying to answer you.

This is the problem the paper GRAVITY tries to solve.

The Problem: The "Messy Attic"

Current AI chatbots that remember long conversations are getting better at finding the right "boxes" (retrieving information). But once they find the box, they hand the AI a messy pile of text. The AI then has to guess how the pieces fit together:

Who is involved? (Relational)
When did it happen? (Temporal)
What is the big story? (Thematic)

Because the AI has to guess these connections on the fly, it often gets confused, especially in long, complex conversations.

The Solution: GRAVITY (The "Organizer")

The authors created a tool called GRAVITY. Think of GRAVITY not as a new brain, but as a super-organized librarian who works alongside the AI.

GRAVITY doesn't change how the AI thinks or how it searches the attic. Instead, before the AI tries to answer your question, GRAVITY takes the messy pile of papers and organizes them into three neat, labeled folders. It then hands these folders to the AI along with the messy papers.

Here are the three folders GRAVITY creates:

The "Who's Who" Folder (Entity Anchors):
Instead of just seeing the name "Caroline" scattered across different notes, GRAVITY creates a profile card for Caroline. It lists her job, her projects, and her relationships. It's like having a contact card that says: "Caroline is a developer working on MedLLM." Now the AI doesn't have to guess who she is; the card tells it immediately.
The "Timeline" Folder (Event Anchors):
Instead of reading a paragraph to figure out that you noticed a problem yesterday and fixed it today, GRAVITY writes down a clear timeline: "Yesterday: Noticed error. Today: Fixed error." It links these events together like beads on a string, so the AI sees the cause-and-effect clearly without having to re-read the whole story.
The "Big Picture" Folder (Topic Anchors):
Sometimes you talk about the same thing for weeks. GRAVity writes a short summary of that whole saga. It's like a chapter summary in a book. If you ask about a project that spanned three months, the AI can read the "Chapter Summary" to get the main idea instantly, rather than digging through 50 individual emails.

How It Works (The Magic Trick)

The best part about GRAVITY is that it's plug-and-play.

It doesn't require you to rebuild the AI's brain.
It doesn't require you to change the AI's software.
It simply takes the raw conversation, organizes it into those three neat folders, and pastes them into the AI's "prompt" (the instructions it reads before answering).

It's like giving a chef a recipe (the messy notes) and a pre-chopped, organized ingredient tray (the GRAVITY folders). The chef (the AI) still cooks the meal, but because the ingredients are prepped, they can cook much faster and make fewer mistakes.

What the Paper Found

The researchers tested this "organizer" on five different types of AI memory systems. Here is what happened:

It works for everyone: Whether the AI was using a simple list, a complex graph, or a compressed summary, adding the GRAVITY folders made it smarter.
The bigger the mess, the bigger the help: The AI systems that were already pretty good got a small boost (about 4-5% better). The systems that were struggling with long conversations got a huge boost (up to 12% better).
It's the organization, not the extra words: The researchers proved that the improvement didn't come just because they gave the AI more text to read. It came because the text was structured. When they gave the AI a messy summary of the same length, it didn't help much. But when they gave it the organized folders, the AI's reasoning improved significantly.

The Bottom Line

GRAVITY shows that for AI to be a good conversational partner over a long time, it doesn't just need to find the right memories; it needs to see them organized. By adding a layer of structure (Who, When, and What) to the conversation, we can help AI reason much better without needing to rebuild the AI itself. It turns a messy pile of notes into a clear, easy-to-read story.

Technical Summary: GRAVITY

Problem Statement

Long-horizon conversational agents rely on memory systems to sustain coherent dialogue across hundreds of sessions. While recent advancements have improved retrieval mechanisms (e.g., dense vector similarity, graph traversal, hierarchical compression), a critical "reasoning gap" persists between retrieval and generation.

Current systems typically feed retrieved fragments to the language model (LLM) as unstructured text. Even when relevant evidence is successfully retrieved, the generator lacks explicit relational, temporal, and thematic structures connecting these fragments. Consequently, the model must implicitly reconstruct cross-fragment connections (e.g., entity relationships, causal event chains, or topic arcs) from flat text. Empirical evidence shows that even with perfect retrieval (an "oracle" setting where all ground-truth evidence is present), accuracy on complex reasoning tasks drops significantly (e.g., from 84.9% to 75.6% when evidence is scattered among distractors). The bottleneck is not the absence of evidence, but the absence of explicit structure to organize that evidence for reasoning.

Methodology: GRAVITY

The authors propose GRAVITY (Generation-time Relational Anchoring Via Injected Topological MemorY), a plug-and-play, architecture-agnostic structured memory module. GRAVITY operates externally to the host memory system, requiring no modifications to the host's retrieval pipeline, embedding models, or internal storage.

Core Design Principles

GRAVITY addresses the three inherent structural dimensions of long-horizon dialogue that raw text fragments fail to expose:

Relational Structure (R): Cross-fragment entity links (e.g., who is connected to what).
Temporal Structure (E): Causal and chronological dependencies between events.
Thematic Structure (T): Macro-level topic arcs spanning multiple sessions.

Two-Phase Workflow

1. Build Phase (Offline)
GRAVITY processes raw conversation utterances ( $U$ ) to extract three complementary knowledge representations, independent of the host's internal data structures:

Entity Anchors: Dynamic profiles for every named entity containing attributes, typed relations, status transitions, and co-occurrence counts. Profiles are updated incrementally with confidence-weighted merging to handle state changes over time.
Event Anchors: Structured tuples in a canonical 4W1O format (Who, What, When, Where, Outcome). Events are linked into chronological "traces" based on shared participants or topics to preserve causal chains.
Topic Anchors: Cross-session summaries that aggregate utterances discussing the same subject into a single narrative arc, capturing macro-level evolution that single-session summaries miss.

2. Inference Phase (Online)
At query time, GRAVITY operates alongside the host's retrieval pipeline:

Anchor Retrieval: It independently retrieves top- $K$ relevant anchors from its knowledge bases using embedding-based reranking. A temporal preservation mechanism ensures temporally relevant events are retained even if their lexical similarity to the query is low.
Query Expansion: Anchor modules generate expanded retrieval queries (e.g., combining entity names with attributes) to broaden the host's search scope.
Context Injection: The selected anchors are formatted into structured blocks (Topic Summaries, Entity Profiles, Event Records) and injected into the host's generation prompt. The prompt instructs the LLM to treat retrieved memories as the primary source of truth while using anchors for disambiguation and gap-filling.

Key Contributions

Structural Diagnosis: The paper identifies and empirically validates that the reasoning gap in long-horizon memory is caused by a lack of explicit structural organization in the generation context, not merely retrieval failure.
GRAVITY Module: A fully portable, architecture-agnostic module that injects structured knowledge (Entity, Event, Topic) into the generation prompt without altering the host system's internals.
Systematic Evaluation: Comprehensive cross-architecture evaluation across five diverse memory systems (ranging from vector stores to temporal knowledge graphs) and two benchmarks (LoCoMo, LongMemEval).

Experimental Results

GRAVITY was evaluated on five state-of-the-art memory systems (LightMem, A-Mem, Mem0, LiCoMemory, ZEP) using GPT-4o-mini as the backbone.

Performance Gains: GRAVITY consistently improved LLM-judge accuracy across all systems and benchmarks.
- LongMemEval: +9.2% (Micro) and +10.1% (Macro) average improvement.
- LoCoMo: +7.5% average improvement.
- Range: Improvements ranged from +3.8% (on the strongest baseline, LightMem) to +12.2% (on the weakest baseline, ZEP).
Inverse Correlation: Gains are inversely correlated with baseline strength. Systems with weaker internal structural organization benefit more, suggesting GRAVITY fills specific structural blind spots unique to each host.
Ablation Studies:
- Structure vs. Volume: Replacing structured anchors with unstructured free-form summaries yielded significantly lower gains (+1.3% vs. +5.7% on LoCoMo), proving the benefit comes from the schema (Entity/Event/Topic) rather than additional context volume.
- Component Necessity: All three anchor types (Entity, Event, Topic) contribute non-redundantly. Removing any single type degrades performance.
- Reranking: Embedding-based reranking is critical; removing it erases most gains, confirming that quality filtering is more important than raw context injection.
Oracle Experiments: Even when all ground-truth evidence is present in the context, adding structured anchors improves accuracy (e.g., from 75.6% to 78.5% in a "scattered" oracle setting), demonstrating that organizational context compensates for imperfect retrieval and aids reasoning.

Significance and Claims

The paper claims that structured context anchoring is a broadly effective, model-agnostic augmentation paradigm for long-horizon conversational memory.

Architecture-Agnosticism: GRAVITY demonstrates that structural deficits in reasoning can be addressed externally. It does not compete with or replace existing memory backbones but augments them via the prompt interface.
Orthogonal Patching: The gains are not due to a fixed pool of redundant evidence. Cross-system analysis shows that the questions improved by GRAVITY are largely disjoint across different hosts (Jaccard similarity of 0.09–0.17), indicating that GRAVITY patches specific structural blind spots inherent to each host's retrieval strategy.
Diminishing Returns: The paper provides a theoretical derivation showing that the magnitude of improvement is inversely proportional to the baseline's structural density. As a host system's internal organization improves, the marginal gain from external anchoring decreases, but the mechanism remains valid across all levels of baseline strength.

In conclusion, GRAVITY establishes that synthesizing scattered evidence into a coherent, query-relevant context through explicit relational, temporal, and thematic anchoring is a critical step toward robust, reasoning-capable long-term conversational AI.