Research map / Large language models / Retrieval and knowledge
Peripheral Memory for LLMs: Integration of Sequential Memory Banks with Adaptive Querying
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes peripheral memory, a RAM-inspired, decoupled memory of sequential KAN-based memory banks with a confidence bank, queried from LLM hidden states, and evaluates it on consecutive knowledge-based model editing.
Problem
Integrating new long-term knowledge into LLMs without harming original performance; existing working, implicit, and explicit memory methods are tightly coupled to the LLM and lack scalability, reusability, and configurability.
Approach
Treat memory as an independent peripheral module (like RAM) that the LLM (like a CPU) reads and writes via its hidden-state query; memory is a chain of KAN-modeled banks passing outputs sequentially, with a final confidence bank producing a sigmoid mask that gates the retrieved feature inserted as a prefix.
Key results
- Under 3K consecutive edits on Llama 3 (8B), the method with 1K archive reaches ZsRE score 0.8643 and COUNTERFACT score 0.7674, versus 0.5543 for WISE and 0.3573 for GRACE on ZsRE.
- Locality is 1.0 on both datasets for the proposed method, since the memory can be unplugged to restore the original model.
- Memory written on Llama3 (8B) can be shared read-only with Gemma2-it (2B) and Phi3 (3.8B), with the shared-memory ZsRE result for Phi3 dropping to 0.0 efficacy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Peripheral Memory with a 1K archive instead of IKE | GPT-J (6B) · ZSRE, 3K consecutive edits | Score: 0.8212 -> 0.8975 | Helped |
| use Peripheral Memory with a 1K archive instead of IKE | GPT-J (6B) · COUNTERFACT, 3K consecutive edits | Score: 0.6839 -> 0.7416 | Helped |
| use Peripheral Memory with a 1K archive instead of WISE | Llama 3 (8B) · ZSRE, 3K consecutive edits | Score: 0.5543 -> 0.8643 | Helped |
| use Peripheral Memory with a 1K archive instead of WISE | Llama 3 (8B) · COUNTERFACT, 3K consecutive edits | Score: 0.4048 -> 0.7674 | Helped |
| add a 1K archive (ablation) | Peripheral Memory on GPT-J (6B) · ZSRE, 3K consecutive edits | Score: 0.8851 -> 0.8975 | Helped |
| add a 1K archive (ablation) | Peripheral Memory on GPT-J (6B) · COUNTERFACT, 3K consecutive edits | Score: 0.6962 -> 0.7416 | Helped |
| add a 1K archive (ablation) | Peripheral Memory on Llama 3 (8B) · ZSRE, 3K consecutive edits | Score: 0.8405 -> 0.8643 | Helped |
| add a 1K archive (ablation) | Peripheral Memory on Llama 3 (8B) · COUNTERFACT, 3K consecutive edits | Score: 0.7069 -> 0.7674 | Helped |
| increase memory bandwidth from 512 to 1024 banks (ablation) | Peripheral Memory on Gemma2-it (2B) · ZSRE, 1K consecutive edits | Accuracy: 0.99 -> 0.98 | Hurt |
| increase memory depth (ablation) | Peripheral Memory on Gemma2-it (2B) · ZSRE, 1K consecutive edits | Accuracy | Helped |
| use query features from earlier layers instead of the 24-th layer (ablation) | Peripheral Memory on Llama 3 (8B) · ZSRE and COUNTERFACT, 3K consecutive edits | Accuracy and Generality | Hurt |
| use query features from very late layers instead of the 24-th layer (ablation) | Peripheral Memory on Llama 3 (8B) · ZSRE, 3K consecutive edits | Accuracy | Hurt |
Surprising findings
- Memory shared to Phi3 gives zero efficacy on ZsRE (score 0.3333), which the authors attribute to ZsRE relying on the LLM's internal knowledge.
- Localization-based editors such as ROME, MEMIT, AlphaEdit and PMET collapse to near-zero scores after consecutive edits.
Limitations
- Memory is queried directly by the last token hidden state, making retrieval sensitive to minor input variations and reducing generalization to semantically equivalent queries.
- Generality drops as stored data grows (from 0.72 at 0.5K to 0.65 at 10K), showing a trade-off between utilization and semantic discriminability.
- Shared-memory transfer failed for Phi3 on ZsRE, so reusability is not uniform across models.
Experimental setup
Datasets: ZsRE COUNTERFACT
Benchmarks: ZsRE COUNTERFACT
Metrics: Efficacy Generality Locality Score Effective Storage Density write latency
Baselines: FT-L LoRA ROME R-ROME MEMIT AlphaEdit PMET KN EMMET GRACE IKE WISE
Builds on: Kolmogorov-Arnold Networks EasyEdit ROME MEMIT
Closest papers by meaning
- WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models (NeurIPS 2024 · similarity 0.89)
- MEMORYLLM: Towards Self-Updatable Large Language Models (ICML 2024 · similarity 0.89)
- A-Mem: Agentic Memory for LLM Agents (NeurIPS 2025 · similarity 0.89)
- M+: Extending MemoryLLM with Scalable Long-Term Memory (ICML 2025 · similarity 0.88)
- Memory Decoder: A Pretrained, Plug-and-Play Memory for Large Language Models (NeurIPS 2025 · similarity 0.88)
- CMT: A Memory Compression Method for Continual Knowledge Learning of Large Language Models (AAAI 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.