Can an AI keep track of what's done and what's pending when you change topics mid-conversation, without starting over?
Can task state persist across a changing conversation without explicit resets?
This explores whether what a conversation is trying to get done (what's been asked, what's finished, what's still pending) can survive topic changes, interruptions and detours, without wiping the slate and starting over.
This explores whether the state of a task can survive topic changes, interruptions and detours without a reset. The corpus says yes, but mostly by moving the state out of the chat transcript instead of trusting the model to remember it. The sharpest evidence is Can task state management alone improve long-horizon agent performance?. It keeps task state in a separate record that is audited against the environment, not against the executor's own claims of progress. That lifted Qwen from 51.8% to 80.7% on WeaveBench. Long tasks seem to fail less from weak reasoning than from having no trustworthy record of where things stand.
A second route is to make the conversation itself the record, so interruptions never break anything. Can frontends handle delegation while staying conversationally engaged? puts delegated requests, their results and the chit-chat in between on one ordered timeline. One loop keeps talking while another runs background work, and results are folded back in when they land. Nothing needs resetting because nothing ever leaves the timeline. The related failure is in Why do dialogue systems lose context when topics return?. Stack-based dialogue managers lose context when a topic they already popped comes back, while attention can reach any earlier turn. A rigid stack is effectively a reset that happens without anyone asking for it.
Both routes raise the question of what to keep and at what scale. How should agent memory split across time scales? splits working memory into dialogue-level pieces (conversation history, scratchpad) and turn-level pieces (examples, task trajectory). Each piece fails differently and needs its own update policy, so 'task state' isn't one thing. reasoning-modeled-as-recursive-subtask-trees-with-kv-pruning-enables-unlim pushes this further. It structures work as a tree of subtasks and prunes the cache by rule, so finished branches drop away while live ones stay. It reasons accurately even while manipulating 90% of the cache. The shape of the task decides what gets forgotten. Can splitting adaptation into two channels reduce forgetting? is a cousin of the same idea. Fast-changing, task-specific lessons go in easily rewritten text context, not in weights, which cuts forgetting.
The catch is that humans do this without any machinery. Why don't language models develop conversation maintenance skills? describes how people keep threads alive through reference repair and topic hand-off, which is relational work that models don't pick up because training rewards predicting information. Can dialogue systems track both speakers' beliefs across turns? adds a related gap: tracking what both speakers believe as understanding moves from partial to shared, which token-level LLMs lack. So today, keeping task state across a changing conversation is engineered scaffolding around the model, not something it does on its own. The corpus has no head-to-head study of reset versus no reset. It only has these pieces, which point the same way.
Sources 8 notes
Separating task state management from execution, using independent environment audits instead of trusting executor claims, improved Qwen 3.7-Plus from 51.8% to 80.7% on WeaveBench. The same model-harness pair showed consistent gains across multiple benchmarks and task types.
Realtime-Venus demonstrates that delegated requests, results, and intervening dialogue can share one ordered record, letting foreground interaction continue while background tasks execute. A dual-loop runtime keeps conversation flowing and folds results back in naturally.
Research shows stack-based dialogue structures lose context when popped topics are revisited, while transformer attention enables systems to retrieve any previous turn without structural loss. Attention-based approaches naturally support the interleaved, revisiting nature of human conversation.
RAISE shows that agent memory consists of four components organized by two design axes: dialogue-level (conversation history, scratchpad) versus turn-level (examples, task trajectory). This granularity distinction predicts different failure modes and update policies for each component.
The Thread Inference Model demonstrates that reasoning structured as recursive subtask trees with rule-based KV cache pruning sustains accurate reasoning beyond context limits, even when manipulating 90% of the cache. This enables single models to replace multi-agent systems by handling full recursive reasoning internally.
Show all 8 sources
Fast-Slow Training routes task-specific lessons into optimized prompts while keeping parameter updates minimal, reaching equivalent performance 1.4–3x faster with substantially less catastrophic forgetting and plasticity loss, demonstrating that forgetting is a misallocation problem rather than an inherent cost.
Humans keep conversations smooth through implicit techniques like reference repair and topic hand-off that sustain relational interaction, not convey information. Language models don't develop these because training signals reward information prediction, not relational work.
CRSA integrates rate-distortion theory with RSA to enable bidirectional belief tracking across dialogue turns. Demonstrated on referential games and doctor-patient dialogues, it captures progression from partial to shared understanding, providing the information-theoretic framework that token-level LLM systems lack.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation
- Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory
- Conversational Alignment with Artificial Intelligence in Context
- LLMs Get Lost In Multi-Turn Conversation
- Collaborative Rational Speech Act: Pragmatic Reasoning for Multi-Turn Dialog
- Dialogue Transformers
- Learning, Fast and Slow: Towards LLMs That Adapt Continually
- PGMem: Tightly Coupled Persona-Memory Graph for Lifelong Personalized Agents