Opening — Why this matters now
LLM agents are getting longer memories, better tools, and more elaborate planning stacks—yet they still suffer from a strangely human flaw: emotional whiplash. An agent that sounds empathetic at turn 5 can become oddly cold at turn 7, then conciliatory again by turn 9. For applications that rely on trust, continuity, or persuasion—mental health tools, tutors, social robots—this instability is not a cosmetic issue. It’s a structural one.
The paper behind today’s discussion makes a quiet but important claim: the problem isn’t missing memory. It’s missing inertia.
Background — Memory is not momentum
Most prior attempts to stabilize agent behavior focus on what the agent remembers: persona prompts, conversation buffers, retrieval-augmented memory, hierarchical context windows. These approaches help agents recall facts and prior statements, but they say nothing about how internal states should evolve over time.
Humans don’t instantly flip emotional polarity when a single sentence changes tone. Emotional states accumulate, resist sudden reversals, and recover gradually. This temporal smoothness—well-studied in psychology and dynamical systems—is almost entirely absent from standard LLM agent architectures.
The result is a class of agents that are context-aware but temporally brittle.
Analysis — Adding physics to feelings
The paper proposes a deliberately minimalist fix: introduce an external, low-dimensional affective state governed by explicit dynamics—without touching the language model’s parameters.
The core idea
- Maintain a continuous Valence–Arousal–Dominance (VAD) vector outside the LLM.
- At each turn, extract a noisy, memoryless affect signal from the dialogue.
- Update the affective state using first- or second-order dynamical rules.
- Feed the resulting state back into the prompt to gently condition generation.
In short: treat affect like a physical system, not a classification label.
Three regimes compared
| Regime | Memory | Inertia | Behavioral Outcome |
|---|---|---|---|
| Stateless | ❌ | ❌ | Reactive, incoherent |
| First-order | ✅ | ❌ | Smooth but shallow recovery |
| Second-order (moderate μ) | ✅ | ✅ | Stable, delayed, human-like |
| Second-order (high μ) | ✅ | ⚠️ Excessive | Emotionally “stuck” |
The second-order model introduces an affective velocity term—effectively momentum. This creates hysteresis: the path down into negative affect differs from the path back out.
Findings — Stability has a measurable shape
The results are surprisingly clean.
1. Stateless agents don’t recover
Without state persistence, valence oscillates around zero, tracking local sentiment but failing to exhibit any coherent descent–recovery arc. No inertia, no story.
2. Moderate inertia enables recovery
With second-order dynamics (μ ≈ 0.8), agents show:
- Delayed emotional descent after adversarial dialogue begins
- Recovery aligned with reconciliation—not before it
- Reproducible affective trajectories across stochastic runs
3. Too much inertia breaks adaptability
High inertia (μ ≈ 0.95) suppresses recovery entirely within the dialogue horizon. Stability turns into rigidity—a familiar control-systems trade-off.
4. Hysteresis scales with momentum
| Condition | Hysteresis (AUC) |
|---|---|
| Stateless | 0.0 |
| First-order | 4.36 |
| Second-order (μ=0.8) | 6.85 |
| Second-order (μ=0.95) | 21.19 |
This monotonic relationship makes inertia a tunable design parameter, not a side effect.
Implications — Agent design beyond prompting
This paper quietly reframes how we should think about long-horizon agent stability.
Key insight: behavioral coherence does not require deeper models, longer context windows, or emotional fine-tuning. It requires explicit state evolution.
Why this matters:
- The mechanism is model-agnostic and inference-time only
- Dynamics are interpretable, debuggable, and deterministic
- The approach generalizes beyond affect (e.g., politeness, assertiveness, risk tolerance)
In other words, we can regulate agent behavior the way engineers regulate systems—by shaping trajectories, not micromanaging outputs.
Conclusion — Stability is a design choice
The industry keeps asking how to make agents “more human.” This paper suggests a more precise question: how do we make them temporally coherent?
By adding inertia—not memory—we get agents that feel calmer, more consistent, and more trustworthy over time. Not because they learned better emotions, but because they were finally given rules about how change is allowed to happen.
Cognaptus: Automate the Present, Incubate the Future.