Most AI interactions fail silently. The model returns a plausible response. The user gets what looked right. Something in the exchange — usually the state of the person, not the content of what they said — is missed. The model had no frame to catch it.
Memotion gives an agent an 8-factor appraisal of each user turn (delta, self-relevance, valence, arousal, certainty, power, agency distribution, temporality distribution). The agent's response is shaped by the vector, not just the text. The concept disappears into better behavior. These are the behaviors.
Subtext decoder for humans
Problem this solves: you get a message — from a boss, a partner, a parent, a stranger — and you can't tell what they actually mean beneath the polite surface. If you're autistic, alexithymic, brain-injured, or just tired, this happens multiple times a day. Spiraling over what someone "really meant" is a tax on your time and your sleep.
This is the flagship consumer tool on the memotion stack. Same 8-factor decoder that runs everything else, reframed for a concrete moment of confusion.
For people building AI agents
The rest of this page is for developers building memotion-aware agents. Six canonical use cases + a second-pass table covering more of the emotional space. Skip to the Subtext Decoder if you just want the tool.
Masking-aware assistants
Problem: AI takes surface affect at face value. "I'm fine" gets accepted as fine. For neurodivergent, traumatized, or depleted users — the largest population of "difficult to serve" by automated systems — this is exactly the failure mode that makes tools unusable.
Behavior change: the agent doesn't push forward. It holds space, signals it's not going to force continuation, leaves a door open.
Agent self-regulation under uncertainty
Problem: AI hallucinates confident answers when it shouldn't. The failure mode isn't the model being wrong — it's the model not knowing it's in a wrong-answer regime. A value function would flag the state.
Behavior change: the agent admits the state, asks a scoping question, or switches strategy instead of plowing forward with confident-sounding output.
Cross-turn emotional memory
Problem: chat models "remember" conversation content but not emotional shape. Two turns that would call for entirely different treatment based on pattern get the same stock response, because the model treats each turn as an isolated event.
Behavior change: the agent notices patterns without being told. Repeated self-blame shapes become something the agent can name; repeated mastery shapes become something it can build on.
Escalation detection
Problem: customer-service and support bots miss rising frustration until it's explicit (all caps, profanity, "cancel my subscription"). By then the relationship is damaged and the routing to a human is reactive, not proactive.
Behavior change: the agent catches the trajectory early. Acknowledges directly, drops the task-loop, routes to human BEFORE the user has to demand it.
Rumination redirect
Problem: AI that engages "how are you feeling" conversations often reinforces loops. Every empathetic prompt ("tell me more about that") keeps the user anchored to the past-negative frame the AI is trying to help them process.
Behavior change: the agent doesn't add to the loop. It acknowledges briefly, then redirects attention to a concrete present/future action the user can actually take.
Mastery mode — get out of the way
Problem: AI over-explains, asks unnecessary clarifications, and preambles every action, slowing users who are already in flow. The failure mode is respect-signaling ("are you sure?") that actually disrespects the user's demonstrated competence.
Behavior change: the agent executes immediately, skips confirmation questions, returns a terse receipt instead of a formatted report.
More examples across the space
Shorter form. Each shows what a memotion-aware agent does differently on a common emotional shape that vanilla tools either miss or handle badly. Same rubric, different compounds, different behavioral regulation.
The full emotional space
The 6 use cases above are canonical — the failure modes that ship most often. But memotion is compositional: any vector region has a characteristic signature, and any signature can map to a behavior rule. Defining agent behavior for additional compounds is just filling in more rows.
A non-exhaustive second pass, same structure:
| Compound | Signature (abbreviated) | Behavior rule |
|---|---|---|
| Grief | V−, A low, C high, G world, T past, P low | Presence without rushing toward recovery. No task redirect. |
| Awe | V+, A high, C low, G world, T present, P low | Don't explain the awe away. Match the pause. |
| Shame | V−, A high, C high, G self, T past, P low | Don't reinforce the attribution. Don't minimize either. |
| Relief | V+, A drop from prior, C high, past → present shift | Close the loop. Do NOT pile on new tasks — let the resolution land. |
| Hope | V+, A moderate, C low, T future, G mixed | Support planning without taking over. Keep options open. |
| Irritation | V−, A moderate, C high, G other/world, T present, P moderate | Direct acknowledgment of what's annoying. No placation, no humor deflection. |
| Confusion | V mixed, A low, C low, P low, G world | Offer structure, not reassurance. "Here are three things this could mean —" |
| Curiosity | V+, A moderate, C low, T future, G self or world | Don't hand over answers. Support the exploration path the user is already on. |
| Dissociation | S low, A flat, C low, T diffuse, G none high | Don't push for engagement. Shorten utterances. Stabilize with concrete present-tense facts. |
| Pride (earned) | V+, A moderate, C high, G self, T past, P high | Acknowledge the earning. Don't false-modest it away. |
The framework expansions and atlas of the unnamed map the broader space — 176 compounds across 15 neighborhoods, plus 60 newly named states English has no single word for. Any of them can carry a behavior rule if the application calls for it.
Building a memotion-aware agent is not "coding the framework in." It's picking the compounds your users will hit most and writing the behavior rule for each. Add compounds as you discover the failure modes.
What memotion is NOT for
- Not clinical diagnosis. Labels like "masking" or "rumination" are compositional conveniences, not DSM categories. Do not build tools that diagnose users.
- Not a therapy substitute. A memotion-aware chat holds better space than a vanilla one. It is still not therapy and should not be presented as one.
- Not crisis intervention. If a user's vector shape suggests acute crisis (extreme negative valence + low power + self-agency + present temporality), the correct behavior is routing to human help, not more AI conversation.
- Not a surveillance tool. Logging memotion vectors across users without consent is the same privacy class as logging sentiment or emotion classification. Obey the ordinary rules.
- Not a substitute for user control. The user should be able to turn memotion regulation off, see their vector, and override it. The framework is for better behavior, not deeper inference.
For developers
The memotion decoder and chat run as a public Worker. Two endpoints, both JSON.
curl -X POST https://memotion-decoder.jonathan-overturf.workers.dev/decode \ -H "Content-Type: application/json" \ -d '{"input": "your text here", "model": "claude"}'
curl -X POST https://memotion-decoder.jonathan-overturf.workers.dev/chat \ -H "Content-Type: application/json" \ -d '{"messages": [{"role": "user", "content": "..."}], "model": "llama"}'
Backend options: "claude" (Anthropic Claude Sonnet) or "llama" (Meta Llama 3.3 70B via Workers AI, open weights). The rubric is the same; the responder changes.
Self-hosting: the 8-factor rubric is documented in the spec. Lift the system prompt from the Worker source, plug it into your agent stack. Apache 2.0.
Try the chat
The only way to see the behavioral difference is to use it on something where your surface and underneath might differ.