Keyboard shortcuts

Press or to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Topic 20: Multi-Turn Conversation and Long Context

Every LLM problem, compounded by time. A single-turn system either works or it doesn’t; a conversation degrades, and it degrades in ways that no single-turn evaluation will ever show you.

Start here: MULTI_TURN_DEEP_DIVE.md, read straight through once. INTERVIEW_GRILL.md is active-recall questions for afterwards, and conversation_py.md is a minimal working loop.

What the deep dive covers

Part 1 — long context. What actually degrades as context grows, separating three claims that routinely get conflated: models attending unevenly by position, performance dropping from sheer length even when the answer is trivially findable, and failure to integrate across distant pieces. The gap between advertised and effective context — the single most useful fact in the area, and one that is measured in orders of magnitude rather than percentages. Why needle-in-a-haystack is a bad benchmark and what replaced it. RoPE and why long context is mechanically hard. Attention sinks, KV eviction, and shipped sparse attention. Then the economics: how the KV cache grows over a conversation, how prefix caching works at each provider, and why appending is nearly free while editing history is not.

Part 2 — the system. Memory architectures and what the current crop of memory products can and cannot be shown to do. Context engineering — compaction, sub-agent isolation, just-in-time retrieval. Evaluating a conversation rather than a response. Multi-turn safety, which is a genuinely different problem from single-turn safety. State, sessions, concurrency, and the production shape. Ends with a reference architecture walked end to end and the ten questions interviewers actually ask.

The three findings worth knowing cold

Multi-turn degradation is a variance collapse, not a capability loss. Across 15 models and 200k+ simulated conversations, splitting a task across turns costs about 39% — but the decomposition is that aptitude falls ~16% while unreliability rises ~112%. The model still knows how; it just stops doing it consistently. The mechanism is premature commitment: it guesses an interpretation early, builds on it, and does not recover. Temperature 0 does not fix it.

Compaction destroys constraints, not facts. Summarizing a long session retains only around 17% of standing session constraints — the “always use British spelling,” “never touch prod” instructions that look like boilerplate to a summarizer and are load-bearing to the user. Recency truncation retains none.

The memory-layer benchmarks are not trustworthy. An independent audit of the standard long-conversation benchmark found 6.4% of the answer key wrong and the standard judge accepting 62.81% of deliberately wrong answers. The strongest published memory-layer paper shows plain full-context beating it on accuracy. The honest case for a memory layer is latency and cost, not quality — and being able to say that is worth more in an interview than quoting a vendor number.

How to use it

Every term is defined at first use. Seventeen “Saying it out loud” blocks give you the words in natural speech — cover the block, say your own version, compare. Evidence is marked throughout, and where the field is genuinely contested or a number is vendor-published, the text says so. A benchmark figure quoted as current when it is a year stale is an easy thing to get caught on.

Next