Concept

• Agent Failure Modes That Grow with One Long Context - Agentic laziness, self-preferential bias and goal drift grow the longer an agent stays in one context window, as do one-shotting the whole job and declaring premature completion.
• Chekhov's Gun Retrieval Fallacy - A model tends to treat any retrieved snippet as meaningfully relevant and feels compelled to use it, so irrelevant hits actively mislead the answer.
• Compaction vs Context Reset - Compaction (an in-place summary) preserves continuity but carries pathologies forward, while a reset (a fresh agent plus a structured handoff) gives a clean slate at the cost of handoff fidelity, and which is right depends on the model.
• Context as a Compiled View - Build each call's working context from richer durable state through named, ordered, observable processors, keeping storage separate from presentation, scoping by default and referencing artefacts by name rather than pasting them.
• Context Directness Spectrum - Classify context by how directly it comes from the user — their own words, indirect nearby sources, or boilerplate that only shapes the response.
• Context Gathering Latency Tiers - How urgently a request needs a response bounds how context can be gathered for it — and drives how much of that context should be prepared in advance.
• Context Grounding - Supply the information the model needs in-prompt, and when required restrict answers to that context so parametric knowledge cannot leak through.
• Context Mind Mapping - Vary the target question's individual words and aspects to surface what context would help answer it, before checking what's actually obtainable.
• Context Rot - LLM performance degrades non-uniformly as input length grows, even on simple tasks, with focused prompts of about 300 tokens strongly outperforming full prompts of about 113k tokens on LongMemEval.
• Context Source Proximity and Stability - Sort candidate context sources by how close they are to the application and how often they change, since both predict how hard they are to use.
• Conversation Artifacts - Structured data attached to transcript messages — UI selections, documents, or tool-derived lists — so the agent sees task context without copy/paste.
• Conversational Agent Context - Assemble preamble, prior conversation, and the current exchange — including tool call/response traces — into the transcript the agent sees each turn.
• Dynamic Context Gathering - Runtime collection of per-instance background under latency, preparability, and comparability constraints — usually the bulk of context-design work.
• Dynamic Context Source Discovery - Find what runtime context to gather by mind-mapping the question and ranking sources by application proximity and stability.
• Elastic Snippets - Represent one piece of retrieved content at several sizes so prompt assembly can fit the largest version the remaining budget allows.
• Four Ways Long Contexts Fail - Long contexts fail through poisoning (a hallucination keeps being reused), distraction (repeating history instead of planning), confusion (too many tools) and clash (contradictory information split across turns).
• Fresh-Context State Summary - Rebuild each agent turn from a written environment summary so the only memory of prior steps is what the harness deliberately records.
• Greedy Prompt Assembly - Add elements highest-value-first or remove them lowest-value-first to fit the token budget, choosing direction based on how elements interact.
• Iterative Token-Level Compression - Estimate each segment's token perplexity conditioned on the already- compressed version of prior segments, not the full original text, so earlier compression decisions correctly influence later ones.
• Last-Minute Context Trimming - When assembled context still doesn't fit the token budget, shrink what's already selected — elide less-relevant lines or summarize — rather than dropping snippets outright.
• LLM Application Feedforward Pass - Convert a user's problem into a model-ready prompt in four steps — retrieve context, snippetize it, score and prioritize it, then assemble the prompt.
• Long-Context Binding Constraints - In long-context agent workflows, raw token capacity ceases to be the bottleneck; specification rigor, retrieval policy, and verification mechanisms become binding.
• Lost in the Middle - Long-context performance follows a U-curve: primacy and recency beat the middle, and longer windows do not by themselves fix position sensitivity.
• Main-Context Sections - Split prompt tokens into read-only system instructions, a fixed working context block, and a FIFO message queue headed by a recursive summary.
• Minimal Prompt Crafter - Fill the token budget from the end of the content backward, keeping the most recent elements with no scoring or prioritization needed.
• Next-Turn Salience Promotion - A middleware warning appended after tool output routinely gets ignored on the very next model turn — promote it into a top-of-context reminder before that turn instead of trailing it after the observation.
• Perplexity-Based Prompt Compression - Drop the lowest-perplexity tokens from a prompt using a small reference model, trading human readability for compression ratio since only the target LLM, not a person, needs to parse the result.
• Prompt Assembly Algorithms - Assembling a final prompt is an optimization problem — pick the highest-value elements that fit the token budget and respect dependencies.
• Prompt Compression Budget Allocation - Give a prompt's instruction and question a light compression ratio and its demonstrations a heavier one, selecting which demonstrations survive by perplexity rather than compressing every segment uniformly.
• Prompt Element Dependency - Some prompt elements require another element already be present, and some exclude each other, both of which prompt assembly must be able to express.
• Prompt Element Importance - Rank how crucial each prompt element is to include, separately from its position, using either a numerical score or discrete priority tiers.
• Prompt Element Position - Where a prompt element appears matters — most elements need a specific order, even though skipping an element outright is usually harmless.
• Query-Aware Contextualization - Place the question or lookup key both before and after long retrieved material so mid-context items can attend to the query while encoding.
• Safeguarded Compaction - Compaction should keep full history on disk, never split tool-call and result pairs, audit summary quality before replacing history, and flush notes to memory first, with the compaction prompt tuned for recall before precision.
• Selection-Based vs Generation-Based Prompt Compression - Prefer dropping low-value tokens over asking a model to rewrite a prompt shorter — generation-based compression can't hit a target ratio reliably, costs more, and tends to lose the reasoning path.
• Smart Zone - The smart zone is the portion of a context window (Pocock cites roughly 150k tokens on current frontier models) within which a model still reasons sharply, so workflows should clear or compact at phase boundaries rather than push on degraded.
• Snippet Formatting by Document Type - Format a retrieved data snippet to match the prompt's document type — conversation turn, report section, or serialized structure.
• Snippet Formatting Goals - A well-formatted prompt snippet is modular, natural in its document, brief, and inert — its token count doesn't shift when neighboring text changes.
• Static vs Dynamic Prompt Content - Split always-on task clarification from per-instance context so instructions stay consistent while dynamic material can change every call.
• Structural Context Over Raw Dumps - Conditioning agents on extracted semantic schemas and structured interfaces consistently outperforms dumping raw repository files or flat documents into the context window.
• Topic-Shift Context Truncation - Drop prior conversation once the topic has moved on, using an inactivity threshold or a relevance-judging model to detect the shift.
• Valley of Meh - The early-to-middle prompt region where in-context recency and lost-in-the- middle effects combine so context is used less effectively than at the edges.
• Write, Select, Compress, Isolate - Context management reduces to four moves: write state out, select what to load, compress what is loaded, and isolate work across sub-agents, sandboxes or partitioned state.