Tag
A practitioner discussion exploring whether long-running AI agent failures stem from model capabilities or from the scaffolding around them, highlighting error compounding, context pollution, and weak self-correction as key failure modes.
This paper introduces Trained Agentic Context Management, where a small model (Qwen3.6-35B-A3B) is finetuned over a minimal agentic harness — a tool to call itself and a tool to read arbitrary token ranges — allowing it to match GPT-5.4's 1M-token context performance on long-document benchmarks using only 8K tokens of context.
APDMem introduces a hierarchical long-term memory architecture for LLM assistants that applies agent-controlled progressive disclosure over four memory layers, achieving strong LongMemEval performance while accessing only 8% of conversation data.
研究提出将智能体记忆类比 Git 仓库，通过 commits 与 branches 管理任务版本，使智能体能丢弃过期需求并从正确版本恢复。方法 GitHarness 在全部 30 个测试设置中优于普通续跑，某设置下编码任务 token 消耗减少 73.6% 且得分更高。
A practitioner compares three memory architectures for long-running agents — full conversation history, vector-store-only retrieval, and a hybrid of structured episodic logs plus vector retrieval — concluding the hybrid approach works best and flagging memory conflicts during mid-task belief updates as a key open problem.
This post reflects on how agent development has mainly been an iterative process of patching weaknesses—adding retrieval, state management, retries, permissions, and evaluation—while noting the ongoing tension between offloading capabilities to supporting systems versus absorbing them back into the models themselves.
AutoCompact trains coding agents to decide when to compact context, what working state to preserve, and how to resume, using judge-corrected trajectories for SFT followed by RL with task-success rewards, yielding +9.2% on SWE-bench Verified and +5.0% on SWE-PolyBench Verified.
A practitioner guide to AI PM OS, a set of reusable agent skills that help product managers structure problems, verify outputs, and supervise AI collaborators to produce stronger product judgments instead of blindly shipping AI-generated PRDs.
论文提出 DTOC（动态工具输出压缩），让智能体在每轮动态隐藏或恢复工具输出，以自适应管理上下文，避免不可逆的有损压缩。实验显示该方法能节省 token 与成本并提升解决率，同时提供 OpenCode 参考实现。
A Berkeley paper finds that LLMs often retain outdated preferences or deadlines in context and keep attending to stale values despite knowing the update; providing the current state explicitly (or steering attention to the newest value) dramatically improves performance without retraining.
Matt Pocock shares a 'prompt of the day' that uses a retrospective review of recent coding agent sessions to spot where agents struggle to find information or rely on stale docs, arguing that improving repo navigability is an underrated way to save tokens.
The paper introduces Enterprise Representation Simplification (ERS) and Enterprise Representation Complexity (ERC), a representation-neutral framework for measuring and reducing the structural complexity of enterprise data representations to improve AI reasoning, governance costs, and performance.
Latent Space profiles MIT's Alex Zhang on Recursive Language Models (RLMs), explaining how models can treat their own prompts as external objects, use subagent swarms and context offloading to generalize, and take on big research bets in academia.
Meta presents Context Language Models (CLMs), a paper exploring how effective it is to give language models native write access to manage their own context, with potential benefits for agentic workflows.
A Meta and University of Washington paper introduces Context Language Models (CLMs), where an agent treats its own context as an editable file it manages via Bash. Zero-shot and RL-trained variants show better accuracy with fewer FLOPs on BrowseComp-Plus and multi-agent tasks, plus a Suffix Cache Reuse trick to offset cache invalidation costs.
The post argues that as an agent's tool set grows, direct tool calling degrades, and the more robust fix is letting the model write code in a sandbox to call tools as functions — reducing schema clutter in context and lowering wrong-call rates — at the cost of sandbox security concerns.
Meta released a paper on Context Language Models (CLMs), which let AI agents manage their own context rather than relying on hard-coded cleanup rules, achieving an 11.4% improvement over Codex-style summarization while using 21.5% less compute.
StateTape models a repository as a symbol-level code graph and marks which symbols each write changes, allowing coding agents' contexts to be rewritten based on actual code dependencies rather than textual history, improving resolve rates across six coding agents and three edit-heavy benchmarks.
The tweet discusses a press release highlighting the importance of context in AI, introducing Connect AI Gateway, which routes, governs, and learns from AI requests to create a system of record for company knowledge.
The article explores how splitting tasks across conversation turns reduces agent accuracy and proposes a memory system using context-specific retrieval types to maintain accuracy, with a hands-on example using the open-source Graphiti engine.