Tag
A roundup noting that four agent frameworks (Haystack, OpenAI Agents SDK, Google ADK, OpenClaw) shipped fixes in the same week for context compaction and human approval handling—the key failure points for long-running agents. The author argues compaction is the new retry logic and urges builders to check how frameworks handle full context mid-task.
提出 Source-Judged Compaction (SJC)，一种免训练的后处理方法，通过可校准的判断模型 Jev 为压缩摘要中的每条目评分是否能在环境输出中找到依据，并移除无来源条目，从而提升长时程 LLM 智能体的跨运行可靠性 2.1%–8.9%。
论文提出 Multi-Session Agentic 强化学习设定，并提出 Mosaic 方法，通过层次化优势估计优化会话级贡献与分配，在 BrowseComp 上以更少会话数优于广播式优势方法。
作者将 1982 年 64KB PDP-11 Unix 的 init.d、fork/wait 与 overlay 机制类比到 LLM 编码代理的上下文压缩问题，通过编号状态文件（write-once 锁存、append-only 负记忆、WAL 式回放）让代理在 200k-token 压缩后不丢失关键不变量。
Remory 提出一种神经记忆网络，用有限的软记忆 token 补充文本摘要，使冻结的 LLM 在仅用 5.2% 输入位置的情况下逼近全上下文表现，并在 BrowseComp 和 Terminal-Bench 2.1 等长时程 agent 基准上带来一致提升、显著减少重复工具输出和错误。
REMORY introduces a neural memory network that supplements textual summaries with a bounded sequence of soft memory tokens for context compaction in long-horizon agents, enabling a frozen LLM to approach full-context performance using only 5.2% of input positions with consistent gains on Qwen3.8-27B and GLM-5.3-Flash.
A developer building a Rust agent runtime asks where the boundary should sit between model-driven reasoning and harness-managed reliability — retries, loop detection, context compaction, rollback, and recovery — and seeks practitioner advice on what belongs where.
The article discusses the trade-off between context compaction and prompt caching in production AI agents, proposing subagents as a strategy to maintain cache efficiency and reduce context pollution.
A new paper introduces COMPINT, an evaluation suite showing that context compactors silently drop session constraints, retaining only 17% on average. A simple SC-aware extractor recovers over 90% retention without modifying the compactor or model.
This paper introduces CompInt, an evaluation suite for measuring how well context compaction preserves user-issued session constraints in LLM systems. It finds current compactors retain only 17% of constraints on average and proposes an SC-aware extractor that achieves over 90% retention without modifying the compactor or LLM.
A developer tests a trending GitHub project addressing agent recovery after context compaction, finding that a durable ledger outside the transcript helps but stricter acceptance tests are needed to verify exact delivery steps and user constraints survive.
Introduces parallel context compaction for long-horizon LLM agents, enabling fine-grained control over summary volume and reducing end-to-end latency compared to sequential synchronous compaction across multiple backbone models.