Get Started
Home
Topics
Search
Library
Research questionHow can long-horizon LLM agents preserve answer quality under tight prompt-token budgets?As interactions accumulate, supplying the full history becomes costly or exceeds context limits, while compressing memory can omit information needed for later answers. The central difficulty is retaining useful information within a small prompt budget.
AI
AI Agents
AI Memory
Evaluation & Benchmarks
Inference Optimization
Machine Learning
Natural Language Processing
Latest papersRecent research connected to this question, newest first.Compact-Memory LLM Agents via Online Max-Member Clustering and Atom-Aware PackingApplies to long-horizon LLM agents using compact, streaming memory in roughly the 2k–5k prompt-token regime. Evidence comes from AMA-Bench and RealMem and compares quality–token trade-offs, memory merging, and retrieved-context assembly; the source notes that higher-token baselines can be stronger outside this regime.research paper · Sep 4, 2026Learning What Not to Forget: Long-Horizon Agent Memory from a Few Kilobytes of LearningThe source studies a kilobyte-scale, CPU-only, language-model-free learned scorer with verbatim extraction for agent and conversational memory. Evidence comes from matched-budget agent comparisons, LoCoMo evaluation, and an annotation-free self-supervision study; the reported results do not establish performance across other architectures or workloads.research paper · Sep 3, 2026
Related questions
How can long-horizon LLM agents learn when to group actions without overcommitting?How can LLM agent configurations be jointly optimized when prompts, tools, and parameters interact?How can LLM agents answer recurring questions over unstructured documents without repeatedly rereading them?How can memory-augmented LLM agents coordinate memory construction, retrieval, and repair over long-horizon interactions?