Tone and context
The right memories and the right tone, inside a token budget.
Before your model runs, Humane assembles its prompt from the user's state, relationship, recent conversation, relevant memories and open tasks, scoring and trimming each section to fit a budget. It also sets the tone, empathy, formality and length the reply should take.
Context · user_42
- system_personakept
- behavioral_constraintskept
- relational_context
- recent_conversation
- memories1 dropped over budget
- active_tasks
- time_situation
- response_instructionskept
tone
warm_supportive
modifier
cautious (grudge)
empathy
0.82
formality
0.35
response_length
short
A prompt built for this user, today
Candidate items are scored by similarity, recency, importance and relationship, with weights you can override per tenant. When a section runs over budget, the lowest-scored items go first, and the core sections (persona, behavioural constraints, response instructions) are never dropped.
- Eight sections, each with its own token budget
- Retrieval weights tunable per tenant and testable in experiments
- Token counts and dropped items reported for each prompt
- A tired user's “ok, thanks” gets a short emoji reply, and the model isn't called
Memory · user_42
- strong
A walk after work helps them reset on bad days
memory_type: fact
- moderate
Prefers short replies late at night
memory_type: preference
- weak
Sister visiting next weekend
memory_type: note
Per-tenant vector store · removed by erase_user_data()
Tone follows state
Tone
A named tone for the reply, such as warm_supportive or warm_grounded, chosen from the user's state and message.
Grudge modifier
At a grudge of 0.4 the tone turns cautious, and at 0.7 cautious and defensive, until the frustration fades.
Dependence modifier
During a dependence cooldown, replies take a boundary-respectful tone.
Empathy
Raised automatically for an upset user and lowered when there's no need to over-care.
Formality
From casual to formal, set per tenant and adjustable per turn by policy rules.
Response length
Short when energy is low, longer when the user has the attention for it.
The style Humane chose
"context": {
"tone": "warm_supportive",
"tone_grudge_modifier": "cautious",
"tone_dependence_modifier": "",
"empathy": 0.82,
"formality": 0.35,
"response_length": "short"
}Context questions
Something else? Ask us.
Works well with
Timing engine
Humane recommends how long to wait before each reply, based on the user's mood, energy, the intensity of their message and the pace of the conversation. Every delay is broken into labelled contributors, so you can see and chart why.
ExploreDurable memory
Every turn is stored in a per-tenant vector memory, searchable by meaning. Seed it from your other systems, search it from your own code and erase it when a user asks to be forgotten.
ExploreExperiments
Define an experiment with two variants, each a set of engine settings, and Humane assigns users and applies the right overlay on every call. Results compare the variants with lift and a two-proportion z-test.
ExploreEvery AI should know who it's talking to.
The safety and human-oversight layer for emotional AI: user state across sessions, risk alerts, safety gates and human takeover, with every decision explained.
