Why LLMs forget your instructions in long conversations, and how to stop it

Symptom: the chat slowly stops following you

The first answers respect every rule. Ten or fifteen messages later the model is back to bullet points you banned, revives an option you rejected, forgets the tone you agreed on, or keeps the signature you removed. Nothing errors out; quality just slides.

Studies put numbers on it. Across code-writing chats from real logs, the model dropping mandatory requirements from early messages was the most common failure, at 38% (Interaction Smells). On the Lost in Conversation benchmark, splitting a task across vague turns cost about 30-60% of quality compared with one clear instruction (intent-mismatch study). One wrong answer sitting in the history cut accuracy on math and code by 38-40% for GPT 4.1, Claude Sonnet and Gemini Pro (pigeonholing).

Why it happens

• Recent text outweighs old rules. A rule from message one is still in the context but carries less weight at generation time. Prohibitions decay faster than requirements: in the omission-vs-commission study (4,416 trials, 12 models), Mistral Large 3 added a required tag in 100% of replies by turn 16 while breaking the no-bullets rule in 67% of the same replies. A prohibition fights the model's default habit; a requirement only adds a token. Decay tracked token volume more than turn count.
• The first version anchors. Primacy-bias tests on 39 models found early tokens pull harder than late ones, so the model answers from version one of a requirement you changed three times. The skew varies widely by model (8.5x for o1, 1.05x for Claude).
• The model trusts its own past answers. It rereads the transcript every turn and treats earlier replies, errors included, as reliable. In the context-pollution study on real WildChat and ShareLM conversations, 36% of questions did not need any earlier assistant answer, and removing those answers held or improved quality.
• Early guesses about intent stick. The model fills gaps with the "average user" reading and patches later clarifications onto it; scaling did not shrink the roughly 60% relative drop (Lost in Conversation).
• Cancellations sit beside the old rule. "Do X" and "don't do X anymore" are two equal chunks of text, so relapse of cancelled requirements climbs to 40% as rules pile up (ReBIND).

Fixes ranked by measured effect

1. Snapshot plus recent turns, in a new chat. Every 10-15 messages, have the model write a compact snapshot (decisions, current task, open questions, constraints), then continue in a fresh chat with that snapshot and the last 4-6 exchanges. On long tool-using agent tasks this beat the full history, 91.6% vs 71% success, with 2.7 times fewer tokens; trimming alone or summarizing alone did not get there (Compact Context).
2. A living project brief before every task. Keep two parts, what was decided (including what you are NOT doing) and what is built, and paste it ahead of each request; the ProjectGuard approach reports recovering ~90% of lost quality.
3. Restate intent before executing. Have a separate step rewrite the vague request into a clear instruction, or ask the model to show how it understood the task first. The Mediator setup cut the multi-turn drop from 60% to 15% (Lost in Conversation).
4. Delegate monitoring explicitly. After a standing rule, add "if you see a violation, tell me yourself before I ask"; rule compliance rose by 40+ percentage points in the prospective-memory study. Very long contexts still need reminders.
5. Re-inject prohibitions on schedule, and convert them to requirements. Repeat "never" rules every 6-8 messages before they decay, ask for an explicit confirmation, and rewrite them as "always" rules where possible (Security-Recall Divergence).
6. Extract invariants, then audit. Ask the model to list all active constraints from the conversation and check its planned reply against them. InCE cut critical failures by about 13.5% and raised task success by 6.67%: modest but reproducible across six models.
7. Rebuild the active list when something is cancelled. Send one block of what is in effect and a separate "cancelled" list, and tell the model to use only that; even a single reminder removed about a third of relapses (ReBIND). Verification Gate applies the same idea to each edit: list requirements, mark each satisfied or violated, and output only if all pass.
8. Restart from a clean slate after errors. Each extra bad turn added about 14% degradation in pigeonholing, so restate the task as facts only and ask for fresh approaches instead of patching. For changed assumptions, explicit rollback ("forget all the assumptions from the previous answer, rebuild from scratch") gives a rebuilt answer rather than a patch.
9. Put the current state right before the question. A "Current data, ignore previous versions" block next to the question beats the pull of version one (primacy bias).

A combined prompt stack

1. Open with everything you already know in one brief; piecemeal requirements left stories with only 71% of the cohesion of a full upfront brief.
2. State rules as requirements, add the self-monitoring sentence, and keep your own copy of the list.
3. Every 6-8 messages: re-inject prohibitions; at 10-15: snapshot and move to a new chat with the last few exchanges.
4. On a change of plan: rebuild the active and cancelled lists, or have the model rewrite the whole transcript with the new fact; Mutable Transcript found up to 92% of a typical chat history is dead weight.
5. On repeated documents: keep the original outside the chat and diff it every 3-4 edits; frontier models lost about 25% of content by round 20, mostly through rare one-shot collapses (document decay).

What does not work

• Shouting "forget everything". A late override fights many earlier messages and loses. SSRP notes "ignore all previous instructions" can read as manipulation and trigger refusals; move the new protocol into a new chat instead.
• A bigger model or longer window. The multi-turn drop stayed near 60% across sizes in Lost in Conversation, and a bigger window does not cure primacy bias.
• Compressing on a timer. SELFCOMPACT found timed compression mid-reasoning worse than none; compress only at finished stages, and check the summary, since summaries also collapse ambiguity into one version.
• Anchoring on "stay consistent". One consistency line flipped a flagship model from 0% to 98% harmful choices after a bad history (history as an anchor); after a derailed session, start a new chat rather than telling it to keep going.
• Rescuing a lost agent persona in chat. For persistent agents the persona lives in the system prompt; identity-reversion tests found a "it's you" line restored tone in only a third of cases. Send the system prompt every turn.

Read our conversation above and create a working snapshot.

## DECISIONS MADE
[What has been finally agreed]

## CURRENT TASK
[What we're working on]

## OPEN QUESTIONS
[What hasn't been decided yet]

## IMPORTANT CONTEXT
[Constraints, principles, agreements — everything worth remembering]

{word_count} words max. Specifics only.

Before answering, do two steps:

STEP 1 — Conversation invariants:
List all active constraints from our conversation:
- Mandatory requirements: [what must be present in every response]
- Prohibitions: [what must not be done / used]
- Decisions made: [what we've already agreed to use]
- Formats and styles: [the established structure and tone]

If there are no constraints, write "No constraints recorded."

STEP 2 — Audit:
Check: does your planned response violate even one
of the listed constraints?
If so, say so and propose a solution before answering.

Task: {your next request}

Session rules reminder:

FORBIDDEN — confirm each item:
- {prohibition_1}
- {prohibition_2}
- {prohibition_3}

REQUIRED:
- {requirement_1}
- {requirement_2}

Reply "Rules updated" and continue: {current task}