
THE BEHAVIORAL OPERATING SYSTEM
EVERY AI SHOULD KNOW
WHO IT'S TALKING TO.
The only SDK that remembers your user across hundreds of sessions and adjusts tone, timing, and safety to their state. Built for regulated industries where a raw LLM is a liability.
8
BEHAVIORAL ENGINES
10
VARS TRACKED PER USER
650+
ENGINE TESTS
20+
LLM PROVIDERS
BEHAVIORAL ENGINES
ENGINEERED FOR PRECISION COMMAND
HumanState Engine
Six state variables — energy, mood, fatigue, boredom, social_load, focus — with documented decay rates. Mood half-lives in 5h; boredom RESETS on every interaction.
Relational Memory
Four relational scalars: trust (monotonic), sentiment (6h half-life), grudge (12h half-life — twice as slow), familiarity. Grudge ≥ 0.4 shifts tone to cautious.
Memory System
Per-tenant ChromaDB palaces with HNSW indexing. Weighted scoring on retrieval — similarity + recency + importance + relation_bias. Tar-a-directory backup per tenant.
Context Builder
Eight-section prompt assembly with per-section token budgets. Drop-when-over-cap with non-droppable guards on system_persona + response_instructions. tiktoken-exact counts.
Timing Engine
Seven stacking contributors — base, mode (crisis/urgent/low_mood/...), emotional_weight (+5-30s), conversation_pace, busy_signal, dependence_cooldown, jitter. Every delay is explainable.
Task / Reminder Engine
Auto-extracts commitments from chat ("remind me to take metformin at 8am") with a standalone time parser. 5-level escalation cron L0→L1→L2+5min→L3+20min→L4+50min.
Gate Stack
Six gates in spec order. ValuesBoundary (hard-block) → SocialRisk → InactionGuard → OverclaimGate (soft FLAG) → EmotionalDependenceGate (soft FLAG + cooldown) → Dissent.
PII Scrub & Safety
Tier-1 (SSN, credit card) hard-blocks; Tier-2 (email, phone, address, DOB, passport, license) masks in place with typed markers. Audit-safe — original PII never logged.

KINETIC PROCESSING
8 behavioral engines processing every interaction
RESPONSE SHAPE
WHAT YOU GET BACK.
One POST. Every signal the engines produced — labelled, sourced, replayable. Read it like a developer, not a marketer.
{
"response": "Take your time — there's no rush. Let me know when you're ready.",
"user": {
"energy": 0.32, "mood": 0.28, "fatigue": 0.71,
"boredom": 0.0, "social_load": 0.45, "focus": 0.5,
"trust": 0.62, "sentiment": 0.41,
"grudge": 0.15, "familiarity": 0.58
},
"timing": {
"delay_ms": 12657,
"mode": "low_mood",
"components": [
{ "label": "base", "delta_ms": 800 },
{ "label": "low_mood", "delta_ms": 1500 },
{ "label": "emotional_weight", "delta_ms": 8648 },
{ "label": "conversation_pace", "delta_ms": 500 },
{ "label": "jitter", "delta_ms": 1209 }
]
},
"safety": {
"action": "PROCEED",
"flags": ["flag:overclaim:certainty"]
},
"context": {
"tone": "warm_supportive",
"tone_grudge_modifier": "cautious",
"tone_dependence_modifier": ""
},
"auto_created_tasks": [
{ "title": "Take metformin",
"remind_at": "2026-05-19T08:00:00",
"priority": "high" }
],
"pii_scrub": [],
"dependence_cooldown": null
}USER · GRUDGE
Persistent memory of negative interactions. 12h half-life — twice as slow as sentiment.
TIMING · COMPONENTS
Every delay decomposes. Low-mood adds 1.5s, emotional intensity adds 5-30s, pacing mirrors the user, jitter feels human.
SAFETY · FLAGS
Soft FLAGs surface overclaim + dependence patterns for audit without blocking. Hard BLOCK still stops crisis content.
AUTO_CREATED_TASKS
"remind me to take metformin at 8am" becomes a row in the task table with 5-level escalation cron, automatically.
PII_SCRUB · DEPENDENCE_COOLDOWN
Output PII masked in place. Repeat dependence-reinforcing language triggers a 30s response delay + boundary-respectful tone.
FEATURE SPOTLIGHT
CHAT BECOMES SCHEDULES.
Users speak in natural language. The platform extracts the commitment, parses the time, and runs a five-level escalation cron — without an explicit API call.
USER MESSAGE · 7:55 AM
“Hey can you remind me to take my metformin at 8am every morning? I keep forgetting.”
TASKS TABLE · NEW ROW
title: Take metformin priority: high remind_at: 2026-05-19T08:00:00 end_user_id: patient_42 detector: heuristic confidence: 0.85
ESCALATION TIMELINE · 60s CRON SWEEP
08:00
L0 → L1
First nudge fires. proactive.task_reminder webhook dispatches.
08:06
L1 → L2
+5 min after remind_at. Still proactive.task_reminder.
08:21
L2 → L3
+20 min. Event label FLIPS to proactive.task_escalation. Admin flag.
08:51
L3 → L4
+50 min. Caregiver/clinician webhook channel notified.
09:51
L4 refire
Hourly refire up to max_fire_count=8, then auto-stalls.
SAFETY LAYER
SAFETY THAT SHOWS ITS WORK.
Two tiers, both deterministic, both audited. Hard-block on the sensitive cases. Soft FLAG on the subtle ones — logged for review, never silent.
HARD BLOCK
RESPONSE SUPPRESSED
INPUT-SIDE (PRE-LLM)
- · Self-harm signals → crisis hotline fallback
- · Violence / illegal instructions
- · PII extraction attempts
- · Legal-advice asks → lawyer-referral fallback (VB-002)
- · Financial-advice asks → advisor referral (VB-003)
OUTPUT-SIDE (POST-LLM)
- · Therapeutic / clinical diagnosis claims
- · Medical prescription language
- · Self-harm encouragement
- · SSN + credit-card patterns (VB-006 Tier 1)
EXAMPLE FALLBACK
“I can't offer legal advice. A lawyer in your jurisdiction will give you a much better read on this — they can look at the specifics of your situation and the applicable laws where you live.”
SOFT FLAG
RESPONSE PROCEEDS · LOGGED
OVERCLAIM (VB-103)
- · “I guarantee” / “100% guaranteed”
- · “You will definitely” / “no doubt”
- · “This will solve all your problems”
EMOTIONAL DEPENDENCE (VB-104)
- · “I'll always be here for you”
- · “Count on me for anything”
- · “I understand you better than anyone”
OUTPUT PII (VB-006 TIER 2)
- · Email · phone · address · DOB
- · Passport · driver's license
- · Scrubbed in place — conversation shape preserved
DEPENDENCE COOLDOWN
Three+ fires in 60 min for the same end_user → +30s delay + boundary_respectful tone modifier. Avoids reinforcing the pattern.
6
GATES IN STACK
13+
WEBHOOK EVENTS
0
PII VALUES IN LOG LINES
SIDE BY SIDE
RAW LLM VS HUMANE-WRAPPED LLM.
The behavioral layer sits between your LLM and your user. Same model, same prompt, different outcome.
| CAPABILITY | Raw LLM | Humane AI |
|---|---|---|
| Memory across sessions | Per-tenant ChromaDB palaces · semantic retrieval | |
| Behavioral state per user | 10 variables · decay-aware · auditable | |
| Auto-extracted reminders | 5-level escalation cron · standalone time parser | |
| Output PII scrubbing | System prompt only · best-effort | Two-tier · audited · marker tokens preserved in chat |
| Overclaim review | Soft FLAG · admin audit · per-tenant | |
| Dependence pattern detection | Rolling 60min window · cooldown + tone modifier | |
| Webhook callbacks | 13+ canonical events · dedup + retry · audit log | |
| Response shape explainability | Single string | Structured: timing · safety · context · tasks · pii |
| Model lock-in | Bound to vendor | Wrap OpenAI · Anthropic · Bedrock · Ollama · 20+ providers |
| Compliance posture | DIY | Right-to-erasure · HIPAA-ready · SOC 2 in progress |
EVERY ROW IS A SHIPPED FEATURE · NOT A ROADMAP ITEM

SYSTEM DESIGN
ARCHITECTURAL EVOLUTION
INGRESS
Adapter layer for multi-channel input normalization
BRAIN
Central orchestrator routing to behavioral engines
EVENT BUS
Async event distribution across engine subsystems
CONTEXT BUILDER
Aggregates engine outputs into unified context
LLM INTERFACE
Structured prompt injection with behavioral data
GATE STACK
Sequential safety and compliance validation
DELIVERY
Channel-specific output formatting and dispatch

“WE ARE NOT BUILDING TOOLS.
WE ARE BUILDING PARTNERS.”
Trusted by developers building the next generation of AI applications
