logicspike/docs

Contact Intelligence

Contact Intelligence — Vision

Status: Active Last Updated: 2026-06-29

What Is Contact Intelligence?

Contact Intelligence (CI) is vlozi's per-contact deep memory layer. It is an opt-in service — disabled by default. When a bot owner enables it, CI transforms any vlozi chatbot from a stateless question-answering machine into a personal AI that genuinely knows the person it's talking to.

Without CI: the chat engine remembers the last few messages. Every conversation starts nearly fresh.

With CI: the chat engine knows this person's goals, personality, emotional patterns, who matters to them, what they're afraid of, what motivates them — and it knows how to approach them specifically. This is the difference between a capable assistant and one that actually feels like your assistant.

The Problem CI Solves

A diet coach chatbot that forgets you're lactose intolerant. An AI tutor that explains multiplication for the third time because it doesn't know you're already in calculus. An AI companion that responds the same way whether you're joyful or quietly falling apart.

These are failures of memory and emotional awareness — not of the LLM. The model is capable. The context is missing.

CI's job is to build that context, maintain it across sessions, and serve it back to the chat engine at the right time, in the right form.

Primary Use Cases

CI is designed for long-term personal AI relationships. The harder the use case, the more CI matters.

Use Case What CI Enables
AI companion / girlfriend Knows your mood patterns, remembers what you shared, notices when you go quiet. Makes the AI feel like it cares.
Diet / fitness coach Tracks goals, restrictions, progress, setbacks over months. Adjusts tone when motivation drops.
AI tutor Knows what you've mastered, where you struggle, your learning style. Never over-explains or under-explains.
Customer retention Knows loyal customers by name, notices when they go quiet, reaches out before they leave.
Therapist / mental wellness Tracks emotional arc over weeks. Detects distress trends. Knows when to shift from problem-solving to just being present.

Normal business chatbots (FAQ bots, product support) don't need CI — the chat engine's session window handles those. CI is for use cases where the relationship itself is the product.

Core Principles

These principles are locked. Every architectural decision follows from them.

1. Generic and smart, not domain-hardcoded. CI doesn't know about dieting or tutoring or companionship. It knows how to build a model of a person from what they say, and it lets the LLM figure out what matters in context. No rules to add, no domains to configure.

2. Behavioral guidance, not fact lists. Context is injected as "approach this person with warmth, not intensity — the November wedding is the real motivator, not the numbers" — not "user facts: lactose intolerant, goal = lose 5kg, diet = keto." The first makes the AI feel perceptive. The second makes it feel like a database.

3. Absence beats noise. If nothing relevant to the current message exists in memory, CI injects nothing. Silence is better than irrelevant context. An AI that confidently says the wrong thing — because it was fed a stale memory about a topic unrelated to today's question — is worse than one that starts fresh.

4. Session-end extraction, not per-message. Memory extraction runs once per session close (a single LLM batch call), not after every message. Real-time extraction is rules-based only — zero LLM cost, non-blocking. This makes the per-contact cost under $0.001 per week.

5. Conflict resolution, never accumulation. When new information contradicts existing memory, CI resolves the conflict: update, supersede, or delete — never just accumulate. A memory store that only adds is a garbage collector waiting to fail.

6. Weekly consolidation keeps accuracy over time. Once a week, CI decays stale memories, compresses repeated episodes into patterns, resolves duplicate entities, rolls up the emotional arc, recomputes the relationship trajectory, and regenerates the core behavioral summary. CI gets smarter every week, not noisier.

What Makes CI Unique

These capabilities don't exist in any open-source memory system (Mem0, Zep, MemGPT, LangChain memory):

Emotional arc — not just current mood, but trend over weeks. Is this person improving? Declining? Volatile? The AI's entire approach shifts based on where they are in their emotional trajectory.

Relationship trajectory — stage (new → building → engaged → loyal → at_risk → dormant) combined with direction and velocity. An early intervention window fires when drift is detected before the stage changes — catch it before they leave.

Proactive outreach — CI doesn't just respond to conversations; it initiates them when signals warrant it. It knows the WhatsApp 24-hour window constraint and uses it strategically to keep conversation threads alive.

Crisis detection — two-level system. Chat engine detects real-time keywords (immediate response). CI detects multi-week distress trends (elevated care mode). Neither replaces the other.

Behavioral guidance injection — the context packet tells the AI how to approach this person, not just what they've said. This is the difference between context that makes the AI feel perceptive and context that makes it feel like it's reading from a file.

How CI Fits in Vlozi

CI is a composable layer that plugs into the chat engine. The chat engine never needs to know how memory works — it just calls CI before each response and uses what it gets back.

owner enables CI per bot
         ↓
chat engine asks CI for context before each LLM call
         ↓
CI returns a context packet (behavioral guidance, mood trend, relationship state)
         ↓
chat engine assembles the prompt and calls the LLM
         ↓
session ends → chat engine calls CI to ingest the session
         ↓
CI extracts memories, entities, emotional signal in one LLM batch call
         ↓
weekly consolidation job refines and improves the contact model

Services that don't enable CI: zero overhead, zero cost, pure session window.

V1 Scope

What ships in V1:

  • Three-tier memory (L1 session core + L2 message relevance + L3 archival)
  • Session-end LLM extraction with conflict resolution
  • Entity graph (people, goals, events, motivations, aversions)
  • Emotional arc (per-session snapshots → weekly rollup → trend)
  • Relationship trajectory (stage + direction + velocity + early intervention)
  • Proactive outreach engine (smart retention + owner-configured rules)
  • WhatsApp 24-hour window awareness
  • Multi-channel identity (master contact + channel links)
  • Crisis detection (two-level)
  • Contact correction (real-time synchronous write)
  • Cold start brief (owner-configured, active until session 5)
  • Weekly consolidation job (9-step pipeline)
  • Dashboard: contact list, contact profile, outreach management

V2 Vision

What follows after V1 is proven:

  • Hybrid memory search — vector + keyword re-ranking, not cosine-only
  • Mid-session L1 update — major revelations (name correction, new goal) update core summary immediately, not next week
  • Response feedback loop — memory importance boosted or reduced based on whether the AI's response landed well
  • Cross-contact tenant patterns — what does the AI learn about all contacts of this type that can improve any individual interaction?
  • Flows integration — CI trigger events (contact.at_risk, contact.milestone) become Flow trigger nodes; owners build custom automations
  • Contact-facing privacy portal — the person the AI is talking to can see and correct their own memory profile
Contact Intelligence