This work identifies a class of user-issued instructions, Session Constraints, that are meant to constrain LLM's behavior for the remainder of a session but are silently dropped during compaction, and introduces COMPINT, an evaluation suite that evaluates compactors across three long-context scenarios: multi-turn chat, agentic trajectory, and long-horizon research.
Abstract
When the context window is under pressure, LLM systems compact prior context to continue ongoing tasks. We identify a class of user-issued instructions, Session Constraints (SCs), such as"do not delete any emails until I confirm,"that are meant to constrain LLM's behavior for the remainder of a session but are silently dropped during compaction. To quantify this loss, we introduce COMPINT, an evaluation suite that evaluates compactors across three long-context scenarios: multi-turn chat, agentic trajectory, and long-horizon research. Current compactors retain only 17% of injected SCs on average, and most perform worse than running the same task without compaction. Retention varies sharply with compactor, prompt, context length, SC phrasing, and injection location, showing that the loss is systematic rather than tied to any single setting. We propose an SC-aware extractor that runs alongside the compactor as a plug-and-play module, achieving over 90% retention across all three scenarios without modifying the compactor or LLM. The COMPINT evaluation suite and accompanying implementation are available at https://github.com/ZhiqiEliWang/compaction-integrity.
Results indicate that explicit, address-based recall can improve information retention and serving efficiency relative to the evaluated context-management baselines under the tested settings.
T. Dang, Yuma Ichikawa, Sakina Fatima et al.· arXiv.org· 0 citations
It is argued that context assembly in ContextPipe is structurally isomorphic to query execution in a relational database: both execute under a hard budget, exploit a tiered cache, and leverage statistics.
Peng Xu, Zu-Yu Zhang, Yu-Ze Sun et al.· 0 citations
LLM agents increasingly take on long-running tasks whose history grows far beyond a single model context window. Existing approaches compress earlier interactions or extract selected information into fixed memory representations, committing to what to preserve before future needs are known. We present Scroll, a context manager that treats each agent session as an executable Session Environment. The environment is backed by an append-only Event Log and a sandboxed, persistent Python kernel. The kernel maintains a typed namespace across model calls, allowing tool outputs, retrieved history, and derived state to be bound to variables rather than serialized into the prompt at each call. Model-written code searches, materializes, and transforms session state through exec; only explicitly printed projections enter the model's working view for the next call. Context management thus becomes a programming task that inherits the improving coding abilities of LLMs, while the Event Log preserves lossless historical ground truth. As the working view approaches its budget, stale spans are evicted but remain recoverable: an eviction index keeps compact landmarks tied to exact Event Log addresses, so that the agent navigates directly to evicted regions instead of searching the full log. With Qwen3.8-Max as the backbone, Scroll achieves 94.8% on LongMemEval_S; 73.1% on BEAM_10M, surpassing the best published memory system by 5.1 points; and 86.7% on LOCA_256K, exceeding the best published long-horizon agent by 37.4 points.
A controlled runtime measurement protocol for reacquisition cost in a bounded-horizon tool-using agent and three models across two task regimes are evaluated, showing that the reacquisition signature is environment-dependent rather than intrinsic to shortening context.
Long-horizon LLM agents rely on context compression to keep growing histories within a finite window [1][2]. These mechanisms are goal-agnostic—retention follows volume, recurrence, and generic salience rather than task relevance—and we show they systematically amplify off-task content, inducing Compression-Induced Goal Drift (CIGD). We formalize CIGD via the drift amplification ratio (DAR), evaluate six compression strategies on a controlled injection benchmark ( per cell), and propose Goal-Anchored Compression (GAC): a pinned goal anchor, a negation ledger, status-aware retention scoring, and drift re-anchoring. Volume- and recurrence-based compressors amplify detours monotonically with repetition (attention-scored eviction reaches at twelve-fold repetition), while query-conditioned compressors de-amplify. Under severe budgets, summarization and scored eviction lose the user’s negated constraint in and of episodes where GAC retains it in ; GAC eliminates residual detour content () with smaller contexts. Prior work asks what compression forgets [3][4]; we ask what it amplifies.
Zhan Zhang, Wen-Zhi Zhang· Asia Pacific Economic and Ma...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
