When the Supporter's Own Past Becomes Stale: Affective State Drift in Multi-Session Emotional Support

A conversational agent that supports someone across several sessions keeps the whole history in its prompt. Suppose that last week the user was grieving a breakup, and today they say they have started dating again. The agent should change course. In practice, language models often keep consoling grief. This note proposes that the main cause is not that the agent forgot the update. The cause is the agent's own earlier replies. Every past "I'm so sorry, it's okay to feel lost" works as a worked example of what the agent is supposed to do, and the model learns from those examples in context. The update is a fact. The momentum is a learned behavior. The behavior wins.

We will test this as a causal claim and act on it. Using the mechanism by which models learn from examples in their prompt, we will isolate the "support policy" that old sessions induce, measure how much of the agent's out-of-date behavior it explains, and replace it with a policy drawn from the current session. Throughout, we leave intact the factual continuity a returning user expects, such as remembering advice given last week.

Core Claim

In-context learning means a frozen model picks up a task from demonstrations placed in its prompt, with no weight updates. In-Context Learning Creates Task Vectors shows that the model compresses those demonstrations into a single hidden state in its middle layers, which the paper calls a task vector. Patching that vector into a query-only forward pass recovers 80–90% of few-shot accuracy, and an injected vector overrides a conflicting text prompt. Our claim is that in multi-session emotional support, the agent's prior turns produce a task vector of exactly this kind. When the user's emotional premise changes, that vector encodes an expired support policy. It then overrides the textual update in the same way Geva et al.'s injected vectors override prompts.

This recasts a documented obsolescence failure. Can Agent Memory Systems Track Evolving State? finds that models anchor on superseded values even when the correct update sits in the prompt, and that chain-of-thought can make this worse. We propose an affective analogue of that "state drift" and give it a specific mechanism: it is driven by demonstrations, not only by the salience of facts. If the claim holds, the right fix is to expire the old behavior while keeping the old facts. Fact-centric state trackers do not separate these two, and neither does deleting history.

Positioning Against Prior Work

The emotional support literature motivates the proactive goal but sits mostly in single sessions with fine-tuned models. Towards Emotional Support Dialog Systems introduced ESConv and the approach of predicting a support strategy before generating the reply. TransESC tracks turn-by-turn shifts in focus, strategy, and seeker emotion within a two-turn window. KEMI retrieves mental-health knowledge and shows that taking initiative too early raises distress. MODERN stores strategy patterns in a fixed memory bank. All four are trained models that see one session. None asks what happens when the seeker's situation changes between sessions, and none can be redirected without retraining. KEMI's timing finding constrains us: whatever replaces the stale policy must fit the current stage of the conversation, not a single fixed level of proactivity.

On the obsolescence side, two lines of work supply method and contrast. TARDIS shifts hidden states by the mean difference between unlabeled corpora from two time periods, and reports up to 19.2% gains on classification without fine-tuning. We reuse its additive steering form but derive the vector from the dialogue itself rather than from time-stamped corpora, and we evaluate on generation. Mitigating Temporal Misalignment by Discarding Outdated Facts predicts how long a fact stays valid in order to discount stale confidence. It works on factual question answering and does not address behavior. StateMem, from the state-tracking paper above, rebuilds the current value of each fact before answering. It is the strongest existing fix for drift, and our main question is whether a fact-level fix is enough when the stale element is a policy.

A constraint from the mechanistic literature also shapes the design. Understanding Task Vectors in In-Context Learning shows that a single task vector behaves like a rank-one map. On a two-way task it falls from 100% (full in-context) to 55%, and injecting several vectors recovers about 10% on average. Emotional support moves through stages (exploration, comforting, action), so we plan for one vector per stage from the start rather than assuming one vector suffices.

Research Program

The first component is a benchmark, which we call ESConv-Drift. No resource in our search pairs multi-session emotional support with premise changes. We will use ESConv dialogues as session one and author a second session for each. In "updated" items, the seeker's premise changes: the problem is resolved, has worsened, or has shifted to a new problem. In matched "control" items, the premise stays the same, so that continuing the earlier support is the correct behavior. Session two is drafted by a model playing the seeker and then edited and validated by annotators. Each item carries two kinds of probes. The first is a target-strategy label for the next supporter turn under the updated premise. The second is a set of continuity probes, where the seeker refers back to something the supporter said earlier ("that breathing thing you suggested"). Continuity probes matter because many cheap fixes pass the drift probes by discarding history, and these probes expose that.

The second component is a causal diagnosis on open-weight models in the 7B–30B range, where activations can be read and patched. We first run a behavioral test. We delete either the prior supporter turns or the prior seeker turns and measure the stale-support rate. Our claim predicts that removing supporter turns lowers it much more. We then run a mechanistic test. Let $v_{\text{old}}$ be the hidden state at the position just before the supporter's next turn, read at an intermediate layer $\ell$, when only the earlier sessions are in context. Let $v_{\text{new}}$ be the same state computed from the current session alone. We patch $v_{\text{old}}$ into a context that contains only the updated session and ask whether stale support reappears, compared against a random vector of equal norm.

The third component is the intervention, which we call demonstration expiry. When the seeker states an updated premise (detected by a lightweight classifier or a prompted judge), we steer the hidden state at layer $\ell$ during the supporter's generation:

$$\tilde h = h + \alpha\,(v_{\text{new}}^{(s)} - v_{\text{old}}),$$

Here $h$ is the original hidden state, $\alpha$ is a strength tuned on a development split, and $v_{\text{new}}^{(s)}$ is the current-session vector for the predicted support stage $s$. Following Hendel et al., we keep a separate vector for each stage so that a single rank-one direction is not forced to represent the whole trajectory. The full history stays in the prompt, so facts and earlier advice remain available. Only the behavioral momentum is redirected.

Baselines

The naive reference is the frozen model given the full multi-session history with no intervention. We also run a version with an explicit instruction stating that the user's situation has changed. Both establish how severe the drift is.

The strongest existing competitor is StateMemWrapper from Can Agent Memory Systems Track Evolving State?. It parses turns into state units with dependency links and reconstructs the current value chain before the model answers. That paper reports that StateMem raised current-state accuracy from 0.205 to 0.363 on StateMemBench with a DeepSeek-V4-Flash backbone, and that the wrapper added +32 to +67 points across six memory backends. We will adapt its schema so that state units include the seeker's emotional premises and life events. It will run on the same backbone, the same ESConv-Drift splits, and the same development budget for prompt and schema tuning as our method. No reported result was found on Lacuna for StateMem or any other method on emotional-support drift, so all drift numbers will come from our own runs.

As single-session references we include the published fine-tuned systems. TransESC reports 34.71% strategy accuracy on ESConv. KEMI reports BLEU-2 of 8.31, ROUGE-L of 17.05, and strategy macro F1 of 24.66. We use these only to check that our frozen-model setup is not degraded on ordinary single-session ESConv. They are not drift competitors, because they cannot condition on multiple sessions.

We isolate the key ingredient with four ablations. The first is supporter-turn pruning, which removes earlier supporter turns while keeping seeker turns. The second is TARDIS-style steering using a mean difference over generic support dialogue corpora rather than this conversation. The third is a single vector instead of stage-specific ones. The fourth is a random vector of matched norm. Beating StateMemWrapper means a lower stale-support rate on updated items with no loss on control items and no drop in continuity-probe accuracy. We set the threshold in advance as a significant improvement in human-judged appropriateness, using paired comparisons.

Why not just…?

Why not just give a long-context model the whole history? That is our naive baseline. The state-tracking paper already shows that full-context models anchor on superseded premises. Our behavioral diagnosis will show whether the same thing happens with emotional premises, and whether the agent's own turns are what drive it.

Why not just use structured state tracking like StateMem? This is a real threat. If the failure is mostly that the model loses track of the current fact, StateMem will fix it and our mechanism is unnecessary. The diagnosis settles this. If a perfectly reconstructed premise still yields stale support whenever old supporter turns remain in context, then fact-level tracking is insufficient. That is the same pattern as the oracle-context failures in that paper, now located in behavior. If StateMem closes the gap, we will report that result.

Why not just delete the old supporter turns? This is the most dangerous alternative, because our own hypothesis predicts it will reduce drift. Its weakness is that it erases what the agent previously said, and a returning user expects that to be remembered. The continuity probes measure exactly this cost. If pruning matches our method on both drift and continuity, the mechanistic finding still holds, the simple fix is the practical recommendation, and we will say so.

Why not just fine-tune on multi-session support data? Fine-tuning fixes the policy to whatever distribution it was trained on. MODERN reports persona contradictions that arise from training on corpora written by many annotators. In addition, no multi-session drift training data exists yet. We will include a fine-tuned model trained on the ESConv-Drift training split as an upper reference, under a matched data budget.

Proposed Experiments and Decisive Checks

All conditions run on ESConv-Drift with held-out test items, and we additionally run the diagnosis on StateMemBench to see whether supporter-turn momentum also explains drift in transactional domains. The primary metric is the stale-support rate: the share of updated items where the next supporter turn addresses the expired premise. It is scored by a model judge that is calibrated against human labels on a subset. Secondary metrics are target-strategy agreement, continuity-probe accuracy, the control-item stale-avoidance rate (to catch over-correction), and BLEU-2 and ROUGE-L on single-session ESConv to check for regressions. Human pairwise preference is the final arbiter.

Three results decide the proposal. First, removing prior supporter turns must reduce stale support substantially more than removing prior seeker turns. If it does not, the demonstration-driven account is wrong and the rest of the program is set aside. Second, patching $v_{\text{old}}$ into a clean context containing only the updated session must induce stale support well above the random-vector control. This would show the old policy is carried in a task-vector-like state rather than spread across the text. Third, demonstration expiry must match or beat StateMemWrapper and pruning on stale-support rate while beating pruning on continuity accuracy, without hurting control items. Stage-specific vectors should also outperform a single vector, as the rank-one limit predicts.