Pick the most reliable safeguard against summary drift in a long conversation

Attempt it

Summary drift is what happens when you compress a long conversation into a rolling summary and each refresh of that summary takes the previous summary as input instead of the raw transcript. The mechanism is compounding distortion. Each summarization step is approximately correct but loses a little detail and adds a little noise. When the next summarization runs on the noisy output instead of the original, the errors compound. By the tenth refresh the summary has drifted far enough from the original to be wrong in ways that look plausible. The correct answer is option 1: **periodically re-summarize from the original transcript** (or a windowed prefix of it). This breaks the compounding chain. The summary anchors to the source rather than to its own prior versions. The other options sound plausible but do not fix the root cause. - **Bigger model** (option 0): reduces per-step distortion but does not prevent compounding. A more accurate game of telephone still drifts. - **Fixed-length summary** (option 2): forces compression at every step, which makes the problem worse by limiting how much detail the summary can preserve. - **Temperature 0** (option 3): reduces variance per call but the systematic distortion is not from randomness; it is from lossy compression. Lower temperature does not undo lossy.

**The correct answer is option 1: periodically re-summarize from the original transcript.** Summary drift is a compounding-error problem. Each summarization step loses information and introduces small distortions. When the next step takes the previous summary as input, those distortions become part of the new input and compound. The summary chain is a noisy auto-regressive process, and after N steps the accumulated error can be substantial, sometimes enough to flip facts or invent details that were never in the original. The structural fix is to break the chain. Periodically (every K turns, or whenever drift signals appear in evals) re-summarize from the raw transcript or from a windowed prefix of recent verbatim turns. The new summary is grounded in the source, not in its own prior version. The compounding chain is broken. **Why the other options are weaker.** - **Bigger model (option 0):** marginally helps because the per-step distortion is smaller, but it does not prevent compounding. The telephone game is more accurate with smarter players but still drifts. - **Fixed-length summary (option 2):** forces every step to compress to a fixed budget. If the conversation grows past what fits in that budget, the summary must drop content; drift accelerates rather than decelerates. - **Temperature 0 (option 3):** reduces variance per call, which is good for reproducibility but does not address the systematic lossiness of compression. Drift is not a randomness problem. ### Production pattern Most 2026 production agents (LangGraph, LangChain ConversationSummaryBufferMemory, custom loops) refresh on a turn-count or budget trigger and re-summarize from the raw transcript every N refreshes. The cost is one extra summarizer call but the quality lift is substantial on long conversations.

Quick reference

TL;DR

Drift compounds when each summary takes the previous one as input; the structural fix is to periodically re-summarize from the original transcript so distortion does not accumulate.

Memory aid

Sign in for the mnemonic that makes this stick. See Pro pricing.

Easy to grasp

Picture the telephone game where each person whispers what they just heard to the next person. By the tenth retelling the message has drifted into something unrecognizable. Now picture the same game where every third turn the next person reads the original message off a sheet of paper instead of relying on the whispered version. Drift stops compounding because the chain is broken. Summary refresh works the same way: re-summarizing from the raw transcript breaks the chain of summaries of summaries.

Key concepts

Context engineering Agents LLM evaluation

Detailed answer & concept explanation

~6 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.

Sign in for 5 free unlocks/day or go Pro for unlimited. See Pro pricing.

Quick scan · 2–4 min

Everything important, quickly.

3 minutes: the compounding-error mechanism, why incremental-only drifts, why re-anchoring breaks the chain, why the other options fail, the 2026 refresh-policy shape.

Real-world examples

Real products, models, and research that use this idea.

LangChain's ConversationSummaryBufferMemory documents the drift problem and recommends periodic full re-summarization from raw transcript.
LangGraph state checkpointers in 2026 preserve raw conversation history alongside the summary specifically to enable periodic re-anchoring.
Anthropic's Claude Opus 4.7 long-conversation guide explicitly recommends a hybrid refresh policy: incremental summaries with periodic full re-summarization.
Letta's hierarchical memory design treats deeper layers as more vulnerable to drift and rebuilds them from lower-layer state periodically.
Production agent frameworks at scale (AutoGen, CrewAI, OpenAI Assistants in 2026) all surface refresh-policy controls in their memory configurations.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QHow would you detect drift in production before users complain?

Sign in or go Pro to see the approach for this follow-up.

QWhat is the cost of full re-summarization and when is it worth it?

Sign in or go Pro to see the approach for this follow-up.

QHow does hierarchical summarization change the drift surface?

Sign in or go Pro to see the approach for this follow-up.

Don't say this

Red flags and common mistakes that signal junior thinking. Click to expand.

Most common mistake

Trying to fix drift with a bigger model or lower temperature. Both reduce variance per step but neither breaks the compounding chain; you need to re-anchor on the source.

Sign in or go Pro to see all red flags and common mistakes. See Pro pricing.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

Sign in for 5 unlocks/day or go Pro for unlimited revision sheets. See Pro pricing.

References

Primary sources. Skim if you want the original framing.