Fortunate Recall: Ontology-Driven Memory Lifecycle Management for Persistent Coherence in LLMs
Fortunate Recall uses a 10+1 behavioral ontology and lifecycle policies to enhance LLM memory management, achieving 76.9% on LifecycleBench.
Key Findings
Methodology
Fortunate Recall (FR) introduces a 10+1 behavioral ontology to classify personal facts and applies lifecycle policies like differential decay, slot-key supersession, event-time validity, and category-aware routing. FR-Bank, its infrastructure-independent implementation, combines LLM metadata extraction with deterministic scoring functions.
Key Results
- FR-Bank achieved 76.9% on LifecycleBench, outperforming Mem0 (61%), A-MEM (65.3%), and MemoryOS (70.5%).
- On LongMemEval-S, FR-Bank scored 75.2% under the Wu et al. protocol, showing no significant degradation in standard retrieval tasks.
- FR-Bank reduced confabulation rates from Mem0's 45.1% to 22.4% while increasing correct answers from 18.6% to 31.2%.
Significance
This research addresses the critical challenge of memory lifecycle management in LLMs by introducing a behavioral ontology and lifecycle policies. It significantly improves long-term coherence in conversational AI and reduces confabulation, enhancing reliability in real-world applications.
Technical Contribution
Key contributions include: 1) a novel 10+1 behavioral ontology for memory classification, 2) a composable lifecycle policy layer for differential memory management, 3) the LifecycleBench benchmark for temporal reasoning, and 4) FR-Bank, a substrate-independent implementation demonstrating the method's generalizability.
Novelty
This is the first work to introduce behavioral ontology for LLM memory management, enabling differentiated lifecycle policies. It outperforms existing systems that rely on uniform decay or RL-based operations.
Limitations
- FR-Bank struggles with explicit retraction scenarios, achieving only a 5% retrieval pass rate on AV7 in LifecycleBench.
- The 10+1 behavioral ontology may not be optimal and requires further refinement.
- Dependence on LLMs may limit its applicability in low-resource environments.
Future Work
Future research could focus on optimizing the behavioral ontology, extending lifecycle policies to handle more complex scenarios, and improving efficiency for deployment in resource-constrained settings.
AI Executive Summary
Large Language Models (LLMs) often struggle to maintain coherence in long conversations due to undifferentiated memory management. Existing systems treat all personal facts equally, leading to unbounded memory growth and declining retrieval precision. To address this, the authors propose Fortunate Recall (FR), a novel lifecycle management framework based on a 10+1 behavioral ontology. By classifying facts into behavioral categories and applying tailored lifecycle policies, FR ensures more effective memory management.
FR-Bank, an independent implementation of FR, achieved 76.9% on the LifecycleBench benchmark, significantly outperforming existing systems like Mem0 and A-MEM. On the LongMemEval-S benchmark, FR-Bank scored 75.2% under the canonical Wu et al. protocol, demonstrating that lifecycle policies do not degrade standard retrieval performance. Additionally, FR-Bank reduced confabulation rates by over 50% compared to Mem0 while increasing correct responses.
Despite its strong performance, FR has limitations, such as its lower effectiveness in handling explicit retraction scenarios. Future work will explore optimizing the behavioral ontology, expanding lifecycle policies, and improving efficiency for broader deployment. The authors have released the ontology, benchmark, and code to facilitate further research in this critical area of LLM development.
Deep Analysis
Background
LLMs often fail to maintain coherence in long-term conversations due to undifferentiated memory management. Systems like Mem0, A-MEM, and MemoryOS have made progress in memory storage and retrieval but fail to address issues like outdated information and retrieval precision degradation.
Core Problem
The core challenge lies in lifecycle management for LLM memory systems. Existing systems cannot distinguish between current and outdated facts or handle time-sensitive information effectively, leading to unbounded memory growth and reduced reliability in long-term interactions.
Innovation
FR's key innovation is the introduction of a 10+1 behavioral ontology that classifies personal facts based on their temporal dynamics rather than cognitive type. This enables tailored lifecycle policies, including differential decay rates, slot-key supersession, event-time validity, and category-aware routing, which significantly improve memory management compared to prior methods.
Methodology
- �� FR uses LLMs to extract metadata such as behavioral category, slot key, lifecycle state, and event-time anchor.
- �� Each fact is categorized into one of 10+1 behavioral types, with specific lifecycle policies like decay rates and supersession rules.
- �� Retrieval combines semantic candidates with category-aware candidates, filtered by a lightweight LLM model to top-20, followed by deterministic lifecycle scoring.
- �� The lifecycle layer operates purely on deterministic math, ensuring efficiency and interpretability.
Experiments
Experiments were conducted on LifecycleBench (516 questions) and LongMemEval-S (500 questions). LifecycleBench evaluates temporal reasoning across 40 personas and 9 attack vectors. Ablation studies on FR-Graphiti and transfer tests on the BEAM benchmark further validated the framework.
Results
FR-Bank achieved 76.9% on LifecycleBench, outperforming Mem0 (61%) and A-MEM (65.3%). On LongMemEval-S, it scored 75.2% under the Wu et al. protocol. Confabulation rates dropped from 45.1% to 22.4%, while correct responses increased from 18.6% to 31.2%.
Applications
FR can be applied to virtual assistants, personalized recommendations, and customer service, where long-term conversational coherence is critical. Its behavioral ontology and lifecycle policies enhance reliability and user experience.
Limitations & Outlook
FR struggles with explicit retraction scenarios, achieving only a 5% retrieval pass rate on AV7. The 10+1 ontology may not be optimal, and its reliance on LLMs could limit deployment in low-resource settings.
Plain Language Accessible to non-experts
Imagine a library where every book represents a memory. Traditional LLMs stack all books on one giant shelf, regardless of whether they're outdated. FR acts like a smart librarian, organizing books by type: some are kept forever, some are replaced, and others expire. For example, your favorite food might change, but your name doesn't. By managing the library this way, FR ensures the right book is always easy to find, saving time and avoiding confusion.
ELI14 Explained like you're 14
Think of your brain as a giant notebook where you write everything: your favorite food, your friend's birthday, and last week's movie. Over time, it gets messy! FR is like a super organizer that sorts everything into categories like 'favorites' or 'important dates.' It even deletes old stuff, like foods you no longer like, or plans that already happened. This way, your notebook stays neat, and you can always find what you need!
Glossary
Behavioral Ontology
A framework classifying personal facts into 10+1 behavioral types to guide memory lifecycle management.
Used to determine lifecycle policies like decay rates and retrieval routing.
Lifecycle Policies
Rules for managing memory, including differential decay, supersession, and event-time validity.
Optimizes LLM memory management and reduces confabulation.
Confabulation Rate
The proportion of incorrect or fabricated information generated by a model.
Used to evaluate FR's accuracy in answering queries.
LifecycleBench
A benchmark with 516 questions to evaluate temporal reasoning and lifecycle management in LLMs.
Validates FR's performance improvements.
FR-Bank
An independent implementation of FR combining LLM metadata extraction with deterministic lifecycle policies.
Used in experiments to demonstrate FR's generalizability and efficiency.
Open Questions Unanswered questions from this research
- 1 How can the behavioral ontology be further optimized for precision?
- 2 How can FR be efficiently implemented in low-resource environments?
- 3 Can FR be extended to handle more complex memory scenarios?
Applications
Immediate Applications
Virtual Assistant Optimization
Enhances long-term conversational coherence, reducing errors and improving user experience.
Personalized Recommendations
Dynamically manages user preference data to deliver more accurate recommendations.
Long-term Vision
Intelligent Memory Systems
Develop systems capable of long-term learning and adaptation for complex applications.
Abstract
Current LLM memory systems treat all personal facts identically, so stores grow without bound while retrieval precision degrades. The core challenge is lifecycle management: which memories should persist, which should be replaced, and at what rate, conditioned on the behavioral type of each fact. Fortunate Recall (FR) is a composable policy layer that classifies personal facts into a 10+1 behavioral ontology and applies category-specific lifecycle policies (differential temporal decay, slot-key supersession, event-time validity, and category-aware retrieval routing) as deterministic functions over LLM-extracted metadata. FR-Bank, our infrastructure-independent implementation, reaches a 76.9% pass rate on LifecycleBench, a new 516-question temporal-disambiguation benchmark, ahead of Mem0, A-MEM, Memory-R1, and MemoryOS (61% to 70.5%), and 75.2% on the full LongMemEval-S under the canonical Wu et al. judge protocol, so lifecycle policies impose no measurable cost on standard retrieval. A pre-registered ablation locates the gains: replacing the typed layer with three generic lifecycle primitives leaves correctness statistically unchanged (-1.7pp, 95% CI [-6.0, +2.7]), so the generic lifecycle metadata carries the correctness advantage, while the behavioral ontology carries calibration, halving downstream confabulation (12.0% vs 24.2%, p<0.001). End-to-end, FR-Bank cuts confabulation from Mem0's 45.1% to 22.4% over answered queries and from 32.2% to 13.0% over all queries while answering more of them correctly (31.2% vs 18.6%); the ranking replicates on the open-weight Kimi K2.5. The decomposition transfers to BEAM, an independently built benchmark: 46.8% correct vs Mem0's 32.9% over 280 questions, with the ontology's benefit concentrated in contradiction resolution and saturating near seven policy clusters. The ontology, benchmark, and code are released.