HebbrixBack to Home
Voice Agents

"Hi, I'm calling about my appointment." Your agent already knows which one.

Voice has no room for hesitation. When a caller speaks, your agent has roughly 300 milliseconds before the silence feels awkward. Hebbrix retrieves everything it knows about a caller in under 50, leaving plenty of room for a natural, informed response.

Start BuildingHow Memory Works

The 50ms window

In a phone call, turn-taking happens fast. Here's where memory retrieval fits in the natural rhythm of conversation.

Caller finishes speaking
0ms
Hebbrix memory retrieval
< 50ms
LLM generates response
50 to 200ms
Awkward silence threshold
~300ms

Memory loads before the LLM even starts thinking. Zero added latency to the caller experience.

Anatomy of a voice call with memory

Three phases. Two API calls. Your agent goes from stranger to informed assistant before the caller finishes their first sentence.

Before

Call connects

  • •Caller ID triggers a memory search
  • •Hebbrix returns name, preferences, history
  • •Context injected into the agent's system prompt
  • •All of this happens in under 50ms
During

Conversation flows

  • •Agent speaks with full context from the start
  • •No "can you repeat your account number?"
  • •New facts extracted in real time with infer:true
  • •Knowledge graph connects related entities
After

Memory updates

  • •Call summary stored as a new memory
  • •Preferences and decisions saved to long-term
  • •Next call starts with even more context
  • •Old memories decay naturally over time

What your agent sees when a call connects

Caller context loaded automatically from previous interactions.

Incoming: +1 (503) 555-0147
4 MEMORIES
Caller
David Chen
Last Call
3 days ago
Open Items
Dental cleaning scheduled Thu 2pm. Wants to switch to Dr. Park.
Preferences
Prefers morning appointments. Has dental anxiety, responds well to calm tone.
Graph Connections
David → patient_of → Dr. Martinez. David → prefers → Dr. Park. David → insurance → Delta Dental PPO.

Works with any voice platform

Hebbrix is a REST API. It works with whatever voice stack you use.

VapiBland.aiRetell AIElevenLabsPlay.aiCustom SIP/WebRTC

One search call when the call connects. One store call when it ends. That's the entire integration.

The difference memory makes

"Can I have your account number?"

"Hi David, I see your account right here."

Caller identification

"What was this regarding?"

"Are you calling about your Thursday appointment?"

Open items recalled

"Would you prefer morning or afternoon?"

"I know you prefer mornings. How about 9am?"

Preferences remembered

"Let me transfer you to the right department."

"I see you wanted to switch to Dr. Park. I can handle that now."

History applied

How Hebbrix features map to voice

Each capability solves a specific voice problem.

Sub 50ms searchFast enough for real-time turn-taking. No awkward pauses while the agent thinks.
3-tier memoryRecent call context in short-term. Ongoing relationships in medium. Permanent preferences in long-term.
Knowledge graphConnects callers to their doctors, appointments, insurance plans, and related entities automatically.
Smart ingestionFeed in call transcripts with infer:true. Hebbrix extracts the facts without any parsing code.
Multi-tenant collectionsOne deployment serves all callers. Each person gets isolated memory with no cross-contamination.
Natural decayOutdated scheduling info fades. Core preferences persist. Memory stays relevant without cleanup.

Give your voice agents a memory

Start with the free tier. 1,000 memories, full hybrid search, sub 50ms retrieval. No credit card needed.

Start Building FreeIntegration Guide