TL;DR
MIGRO is a decoupled long-term memory system that allows a user's interaction history to persist and grow even as the underlying Large Language Model (LLM) is upgraded or replaced. By separating exact evidence preservation from approximate semantic indexing and using LLM-based lossless compression, MIGRO achieves state-of-the-art performance while using up to 99.3% less private memory than existing baselines.
What Problem Does This Paper Address?
Most current LLM memory systems are "model-bound"—they are tightly coupled to a specific model's vocabulary, tokenization, or internal latent states. This creates a major bottleneck: when you upgrade your model (e.g., from Qwen-4B to Qwen-32B), you typically lose the accumulated user history or must pay a massive computational cost to rebuild the memory from scratch.
Furthermore, many systems summarize information at "write-time." This is problematic because a summary fixed today might discard details that become crucial for a question asked six months from now. The authors argue that memory should outlive the model, treating user experience as a persistent state that can be handed off between successive reasoning engines.
Methods
MIGRO introduces a decoupled architecture based on two core principles:
- Asymmetric Fidelity with Deferred Selectivity: Instead of summarizing at write-time, MIGRO preserves historical evidence exactly (canonical source-token IDs). It organizes this evidence approximately using a shared semantic index for retrieval. The decision of what is relevant is "deferred" until a specific query is known, ensuring no information is prematurely lost.
- Portable Host Interfaces: To handle different models with varying vocabularies and embedding spaces, MIGRO uses host-specific "transport" interfaces. These interfaces map the stored canonical evidence into inputs the current "host" model can understand at read-time.
- Reversible Causal Coding: To make long-term storage practical, MIGRO uses a frozen LLM as a probability estimator for arithmetic coding. This allows for lossless, high-density compression of the memory state, making it far more compact than raw text or vector databases.
Datasets
The authors evaluated MIGRO across five diverse benchmarks covering different aspects of long-term memory:
- LoCoMo: Conversational recall and cross-session reasoning.
- LongMemEval-S: Multi-session QA and temporal understanding.
- MemOps: Evolving memory operations like updating or forgetting information.
- MemoryAgentBench: Incremental interaction and handling revised knowledge.
- BEAM-1M: Long-horizon recall within massive, million-token interaction histories.
Key Insights and Ideas
- The "Enduring Object": The user's history is the primary asset; the LLM is just a swappable reasoning engine.
- Fidelity Separation: You don't need high precision for the index (where things are) as long as you have perfect precision for the evidence (what things are). Quantizing the index saves massive space without hurting retrieval quality.
- LLMs are World-Class Compressors: By using the statistical regularity of language captured by LLMs, MIGRO can store vast histories at a fraction of the usual cost (averaging ~1.8 bytes per token for million-token histories).
- Continuity of Accumulation: Portability isn't just a one-time transfer; it’s the ability to keep writing to the same inherited state using a new model, maintaining a continuous thread of experience.
Key Results
MIGRO consistently outperformed all seven baselines (including Mem0, TiMem, and LSTMem) across all benchmarks:
- Efficiency vs. Quality: On the LoCoMo benchmark, MIGRO scored 7.3% higher than the strongest baseline while using 99.3% less private memory.
- Seamless Handoff: After switching models, MIGRO retained 99.9% of the quality achieved by a model that had seen the full history from the start.
- Scalable Storage: For million-token histories, MIGRO used 57.5% less storage than the most compact baseline.
- Joint Frontier: MIGRO achieved 2.60x the answer quality per private MB compared to the best-performing baseline.
Limitations and Future Work
While MIGRO excels at fact preservation and retrieval, the authors note that certain complex reasoning tasks remain difficult. Specifically, contradiction resolution and precise event ordering in very long histories (million+ tokens) still pose challenges for current readers.
Future work may involve improving the "discourse links" between records to help models better navigate temporal relationships and conflicting information. Additionally, the current implementation does not yet support crash-consistent transactional updates or incremental neural recoding, which would be necessary for production-grade database integration.
Summary updated