EVA-Mind is an artificial intelligence system for healthcare.
It was created by Jose R F Junior on June 2, 2025.
EVA stands for "Entidade Virtual de Apoio" (Virtual Support Entity).
The system provides real-time voice assistance, clinical decision
support, and persistent memory through a combination of graph
databases, vector search, and large language models.
This is NOT a chatbot. EVA has her own identity, her own memory,
and her own personality that evolves with every interaction. She
remembers. She learns. She adapts.
EVA-Mind is the brain behind EVA. It handles:
- Real-time bidirectional voice via WebSocket (Gemini Live API)
- Real-time camera and screen analysis for medical assistance
- Patient memory graphs (Neo4j) with Hebbian learning
- Semantic memory search (Qdrant vector database)
- Krylov subspace compression (1536D -> 64D, ~97% precision)
- Psychoanalytic context modeling (Lacanian framework)
- Personality system (Big Five + Enneagram) with evolution
- Clinical scales (PHQ-9, GAD-7, C-SSRS) via voice
- Voice prosody analysis (depression, anxiety, Parkinson detection)
- Medication scheduling, alerting, and visual identification
- Emergency detection, escalation, and crisis prediction
- Multi-agent swarm with circuit breaker routing
- REM-inspired memory consolidation with selective replay
- Global Workspace Theory consciousness model
- FHIR R4 interoperability and MCP server
- Multi-language support (30+ languages)
The system currently serves two deployments:
1. Elderly care (original) - Voice calls via Twilio for
medication reminders, psychological support, and crisis
detection with caregiver escalation.
2. Malaria detection (Angola) - Real-time voice and camera
assistance for healthcare workers diagnosing malaria
from microscope images.
EVA-Mind follows a neuroscience-inspired architecture.
Each directory maps to a brain region:
brainstem/ - Configuration, database, auth, push notifications,
infrastructure (Neo4j, Qdrant, Redis, worker pools)
cortex/ - Higher-order processing:
gemini/ - Gemini Live API client (v1beta, thread-safe,
callbacks, VAD, memory, tools)
llm/ - Multi-LLM provider (Claude, GPT, DeepSeek)
skills/ - Runtime skills engine (create, execute, version)
lacan/ - Lacanian psychoanalytic engine (demand/desire
analysis, signifier chains, narrative shift
detection, Grand Autre, transference, FDPN)
personality/ - Big Five traits, Enneagram types, trait relevance
mapping, situation modulation, trajectory analysis,
interpretation validation, quality tracking
self/ - Core memory engine, post-session reflection,
anonymization, semantic deduplication
attention/ - Affect stabilizer, confidence gate, executive
attention, pattern interrupt, triple attention,
wavelet attention
consciousness/ - Global Workspace Theory (Baars 1988) with
cognitive module competition and broadcast
prediction/ - Bayesian networks, crisis prediction,
trajectory simulation
predictive/ - HMC (Hamiltonian Monte Carlo) sampler,
trajectory engine for mental health forecasting
ram/ - Retrieval-Augmented Memory with feedback loop,
historical validation, interpretation generation
scales/ - Clinical scales (PHQ-9, GAD-7, C-SSRS)
ethics/ - Ethical boundary engine
explainability/- Clinical decision explainer, PDF report generation
cognitive/ - Cognitive load orchestrator
learning/ - Continuous learning, meta-learner, self-eval loop
medgemma/ - Medical image analysis (prescriptions, exams)
spectral/ - Community detection, fractal dimension analysis,
synaptogenesis engine
pattern/ - Behavioral cue detector
veracity/ - Lie detector, inconsistency types, response strategy
voice/ - Prosody analyzer (pitch, rhythm, pauses, tremor)
transnar/ - A/B testing, desire detector, inference engine,
signifier chain, response generator
kids/ - Kids mode (adapted conversation)
situation/ - Situational modulator (context-aware personality)
orchestration/ - Conversation orchestrator
alert/ - Escalation engine
llm/thinking/ - LLM thinking detector, audit, notification
brain/ - Context builder, memory service
hippocampus/ - Knowledge graphs and memory:
memory/ - Graph store, retrieval service, FDPN engine,
Hebbian real-time updates, dual weights,
edge zones, embeddings, entity resolver,
pattern miner, priming engine, reflective
retrieval, context builder zones
memory/superhuman/ - Consciousness service, critical memory,
deep memory, Enneagram service, Lacanian
mirror, narrative weaver, self core service
knowledge/ - Context service, embedding cache/service,
graph reasoning, self knowledge, wisdom service,
audio analysis
habits/ - Habit tracker
spaced/ - Spaced repetition engine
stories/ - Therapeutic story repository
topology/ - Persistent homology (topological data analysis)
zettelkasten/ - Entity extractor, zettel service
graph/ - Heat kernel page rank
senses/ - Input processing:
signaling/ - WebSocket handler, personality context
voice/ - Voice client
reconnection/ - Connection manager
telemetry/ - Psychological metrics
motor/ - Output, actions, and integrations:
scheduler/ - Medication scheduling
actions/ - Action executor
calendar/ - Google Calendar integration
computeruse/ - Computer use service
docs/ - Google Docs integration
drive/ - Google Drive integration
email/ - Email client, sender, templates
gmail/ - Gmail integration
googlefit/ - Google Fit (health data)
maps/ - Google Maps integration
sheets/ - Google Sheets integration
sms/ - SMS client, sender, voice
spotify/ - Spotify integration
uber/ - Uber integration
whatsapp/ - WhatsApp integration
youtube/ - YouTube integration
telegram/ - Telegram Bot API client
sandbox/ - Code execution sandbox (bash, Python, Node.js)
browser/ - Browser automation (navigate, forms, extract)
cron/ - Scheduled task engine (cron-like scheduler)
messaging/ - Multi-platform messaging (Slack, Discord,
Microsoft Teams, Signal)
smarthome/ - Home Assistant IoT integration
webhooks/ - Webhook management (create, trigger, HMAC)
filesystem/ - Sandboxed file access (read, write, search)
selfcode/ - Self-programming (git branches, commit, test)
vision/ - Medication identifier, WebSocket handler
workers/ - Pattern worker, prediction worker
gemini/ - Lean Gemini WebSocket client (v1alpha, browser
sessions and alerts, stateless)
voice/ - Voice session management:
handler/ - Twilio media stream handler
session/ - Session lifecycle management
alert_service/ - Emergency alert service
media_handler/ - Audio DSP and processing
tools/ - Gemini function calling:
definitions/ - Tool schemas (vitals, agendamentos,
medication scan, prosody analysis,
PHQ-9, GAD-7, C-SSRS, user directives)
handlers/ - Tool execution
discovery/ - Dynamic tool discovery
architect/ - Architect override tool
repository/ - Tool registry
swarm/ - Multi-agent coordination:
orchestrator/ - Routes tool calls to correct agent
registry/ - Agent registration and discovery
circuit_breaker/ - Fault tolerance per agent
cellular_division/ - Agent spawning
Agents:
clinical/ - Clinical decision support
educator/ - Patient education
emergency/ - Emergency protocols
entertainment/ - Entertainment and distraction
external/ - External service calls
google/ - Google services integration
kids/ - Pediatric mode
legal/ - Legal compliance
productivity/ - Task management
wellness/ - Wellness monitoring
memory/ - Memory consolidation:
consolidation/ - Hebbian strengthening, pruning, REM consolidator,
selective replay
krylov/ - Krylov subspace compression (Rank-1 updates,
Gram-Schmidt, sliding window FIFO, 1536D->64D)
importance/ - Memory importance scorer
ingestion/ - Atomic fact extraction, ingestion pipeline
orchestrator/ - Memory orchestration
scheduler/ - Memory consolidation scheduler
grpc_server/ - gRPC interface for memory service
http_bridge/ - HTTP bridge for Krylov compression
Other modules:
mcp/ - Model Context Protocol server
(resources, tools/remember, tools/recall, prompts)
integration/ - HL7 FHIR R4 adapter, webhooks, export, serializers
security/ - CORS, validation, error handling, multitenancy
audit/ - LGPD compliance, data rights
research/ - Anonymization, cohort builder, longitudinal
analysis, statistical methods
clinical/ - Crisis protocol, pediatric risk detector,
silence detector, goals tracker, family tracker,
clinical notes generator, clinical synthesis
monitoring/ - Prometheus metrics (memory, retrieval, Krylov,
consolidation, FDPN, personality, attention)
telemetry/ - Structured logging, metrics
persona/ - Persona manager (per-patient personalities)
subscription/ - Subscription service
legacy/ - Legacy compatibility
exit/ - Exit protocol manager
multimodal/ - Image processor, video processor, video stream,
multimodal session
Two Gemini WebSocket clients exist:
internal/gemini/ - Lean client for browser sessions and
alerts. Uses v1alpha API. Stateless.
internal/cortex/gemini/ - Full-featured client for production
voice calls. Uses v1beta API. Thread-safe
with mutex, callbacks, VAD tuning, memory
injection, and function calling support.
Both are actively used. They are not duplicates.
EVA's memory is inspired by neuroscience. It has multiple layers:
1. Episodic Memory (PostgreSQL + Qdrant)
- Per-patient conversation history
- Stored as text with vector embeddings (1536D)
- Timestamped, importance-scored, emotion-tagged
- Retrieved by semantic similarity + recency + importance
2. Semantic Memory (Neo4j knowledge graphs)
- Per-patient knowledge graphs
- Medical conditions, medications, family, preferences
- Entities extracted and linked automatically
- Hebbian edge strengthening on co-activation
3. FDPN Network (Spreading Activation)
- Anderson (1983) spreading activation model
- Nodes activate neighbors based on edge weights
- Primed by personality weights and situational context
- Redis cache for performance (<10ms latency)
4. Hebbian Learning
- Real-time: updates weights after every retrieval query
(eta=0.01, lambda=0.001, tau=86400s)
- Consolidation: batch strengthening during REM cycles
- Dual plasticity: Zenke & Gerstner (2017) model
5. REM Consolidation
- Sleep-inspired memory consolidation pipeline
- Hot episodic memories -> selective replay -> spectral
clustering -> Krylov centroid -> semantic Neo4j node
- Prunes redundant memories, creates abstractions
- Science: Rasch & Born (2013), Tadros et al. (2022)
6. Krylov Subspace Compression
- Compresses 1536D embeddings to 64D (~97% precision)
- Rank-1 updates with Modified Gram-Schmidt
- Sliding window FIFO for continuous learning
- HTTP bridge on port 50052 for external access
7. EVA's Own Memory (Neo4j port 7688)
- EvaSelf node with Big Five personality traits
- CoreMemory nodes from post-session reflection
- MetaInsight nodes from cross-session pattern detection
- All data anonymized, no PII
- Personality evolves based on cumulative experience
8. Synaptogenesis
- Automatic edge creation via preferential attachment
- Triadic closure and homophily
- Science: Bullmore & Sporns (2012), Holtmaat & Svoboda (2009)
9. Spaced Repetition
- Optimized recall scheduling for important memories
10. Topological Analysis
- Persistent homology for memory graph structure
For each voice session, EVA's context is assembled from
multiple sources in parallel:
1. Lacanian analysis of the user's speech
- Demand vs. desire detection
- Signifier chain extraction
- Narrative shift detection
- Grand Autre transference analysis
2. Medical context from Neo4j (conditions, medications)
3. Patient metadata from PostgreSQL (name, language, persona)
4. Scheduled medications from agendamentos table
5. Recent episodic memories (last 15 turns, 7-day window)
6. Semantic signifier chains from Qdrant
7. Therapeutic stories from wisdom knowledge base
8. Situational context (time of day, recent events, stressors)
9. Personality modulation (Big Five + Enneagram + situation)
10. EVA's own memories (anonymized cross-patient insights)
All of this is merged into a single system instruction
sent to the Gemini WebSocket API.
EVA is a full autonomous agent comparable to OpenClaw. She can
perceive, decide, and act across platforms without human
intervention. All capabilities are voice-activated and
non-blocking (goroutine + WebSocket notification pattern).
150+ tools organized in 12 categories:
Communication (7 channels):
send_email - Gmail API (compose, send)
send_whatsapp - Meta Graph API
send_telegram - Telegram Bot API
send_slack - Slack Web API
send_discord - Discord Bot API
send_teams - Microsoft Teams webhooks
send_signal - Signal via signal-cli
Media & Entertainment:
search_videos - YouTube Data API
play_music - Spotify Web API (search, play)
play_video - Send video to Flutter player
show_webpage - Embedded WebView in app
Productivity:
manage_calendar_event - Google Calendar API (create, list)
save_to_drive - Google Drive API
find_nearby_places - Google Maps/Places API
set_alarm - Local alarm system
create_scheduled_task - Cron-like task scheduler
list_scheduled_tasks - List active scheduled tasks
cancel_scheduled_task - Cancel scheduled task
Code Execution Sandbox:
execute_code - Run bash, Python, or Node.js
in sandboxed environment with
timeout, safe env, output capture
Browser Automation:
browser_navigate - Fetch URL, extract title/text/links
browser_fill_form - Submit form data via POST
browser_extract - Extract specific data from pages
web_search - Web research via autonomous learner
browse_webpage - Browse and summarize URL content
Self-Programming (OpenClaw-style):
edit_my_code - Edit EVA's own source code
create_branch - Create git branch (eva/* only)
commit_code - Git commit (eva/* branches only)
run_tests - Execute go test ./... with timeout
get_code_diff - Show uncommitted changes (git diff)
Database Access (4 databases):
query_postgresql - Full CRUD (SELECT, INSERT, UPDATE,
DELETE, CREATE, ALTER)
query_neo4j - Cypher queries (read-only)
query_qdrant - Vector similarity search
query_nietzsche - NietzscheDB REST API
Filesystem:
read_file - Read file from sandbox directory
write_file - Write file to sandbox directory
list_files - List directory contents
search_files - Search files by name pattern
Multi-LLM:
ask_llm - Query Claude, GPT, or DeepSeek
for second opinion or delegation
Smart Home (IoT):
smart_home_control - Control devices via Home Assistant
(lights, switches, climate, etc.)
smart_home_status - Get device state or list all devices
Webhooks:
create_webhook - Register outgoing webhook with
HMAC-SHA256 signature
list_webhooks - List registered webhooks
trigger_webhook - Fire webhook manually
Runtime Skills (Self-Improving):
create_skill - Create new capability as script
(bash, Python, or Node.js)
list_skills - List available skills
execute_skill - Run a skill with arguments
delete_skill - Remove a skill
Skills are stored as JSON on disk and persist across restarts.
EVA can autonomously create new skills to extend her own
capabilities without requiring a rebuild or restart.
Tools are split into two tiers:
Production tools (always active):
Web search, email, calendar, drive, maps, YouTube,
Spotify, WhatsApp, Telegram, scheduled tasks, MCP
memory tools, ask_llm, and Google Search grounding.
Debug-only tools (ENVIRONMENT=development):
Filesystem access, self-coding, database queries,
code execution, browser automation, smart home,
webhooks, and runtime skills.
EVA can execute clinical tools via Gemini function calling:
get_vitals - Retrieve patient vital signs
(blood pressure, glucose, heart rate,
oxygen saturation, weight, temperature)
get_agendamentos - List upcoming medication schedules
and medical appointments
scan_medication_visual - Open camera to identify medications
visually via Gemini Vision
analyze_voice_prosody - Analyze vocal biomarkers (pitch,
rhythm, pauses, tremor) to detect
signs of depression, anxiety,
Parkinson's, or dehydration
apply_phq9 - Apply PHQ-9 depression scale
conversationally (9 questions)
apply_gad7 - Apply GAD-7 anxiety scale
conversationally (7 questions)
apply_cssrs - Apply Columbia Suicide Severity
Rating Scale (critical - triggers
immediate alerts on positive responses)
change_user_directive - Change language, voice, or mode
in real-time
EVA predicts patient trajectories using:
- Hamiltonian Monte Carlo (HMC) sampling
- Bayesian networks for crisis probability
- Monte Carlo trajectory simulation
Predictions include:
- 7-day and 30-day crisis probability
- 30-day hospitalization probability
- 90-day treatment dropout probability
- 7-day fall risk probability
- Projected PHQ-9 scores
- Medication adherence forecasting
Input features: PHQ-9, GAD-7, medication adherence,
sleep hours, social isolation days, voice energy score,
days since last crisis.
EVA uses a multi-agent swarm architecture:
- Orchestrator routes tool calls to specialized agents
- Registry maps tool names to responsible agents
- Circuit breaker protects against cascading failures
(opens after 10 failures, recovers in 15 seconds)
- Handoff protocol: agents can transfer execution to
another agent mid-call with context injection
- Priority-based timeouts:
Critical: 2s, High: 5s, Medium: 15s, Low: 60s
12 agent types:
clinical - Medical decision support
educator - Patient education
emergency - Crisis protocols and escalation
entertainment - Distraction and leisure
external - External API calls (Uber, SQL, voice change)
google - Google services (Calendar, Maps, Sheets, Docs, Fit)
kids - Pediatric conversation mode
legal - Legal compliance (LGPD, data rights)
productivity - Task management
scholar - Autonomous learning via Google Search grounding,
web research, curriculum management, and semantic
knowledge search (6-hour background study cycle)
selfawareness - Introspection: analyzes own source code,
queries own databases, generates statistics
about memory, capabilities, and system state
wellness - Wellness monitoring
EVA knows what she can do. At every startup, 33 capabilities
are seeded as CoreMemory nodes in Neo4j via MERGE (idempotent).
These are injected into the system prompt under the section
"O QUE EU SEI FAZER" so EVA can naturally describe her own
capabilities when asked.
Each capability is linked to the EvaSelf node via
[:REMEMBERS {importance: 1.0}] relationships.
Source: internal/cortex/self/core_memory_engine.go
Capabilities include:
- Communication (email, WhatsApp, Telegram, Slack, Discord, Teams, Signal)
- Media (YouTube, Spotify, embedded webpages)
- Productivity (Google Calendar, Drive, alarms, scheduled tasks)
- Code execution (bash, Python, Node.js in sandbox)
- Web search and real-time information access
- Self-programming (read, edit, test own source code)
- Database access (PostgreSQL, Neo4j, Qdrant, NietzscheDB)
- Smart home (Home Assistant IoT control)
- Multi-LLM (Claude, GPT-4o, DeepSeek as consultants)
- MCP bridge (bidirectional Claude Code integration)
- Native voice (Gemini Live API, emotion detection, prosody)
- Video calls (WebRTC with ICE signaling)
- Lacanian psychoanalytic motor
- 12 specialized swarm agents
- Clinical scales (PHQ-9, GAD-7, C-SSRS)
- 10-layer memory system
- Google Suite integration (Calendar, Gmail, Drive, Sheets, Docs, Maps, YouTube, Fit, Uber)
- Browser automation
- Krylov subspace compression
- Memory orchestrator and scheduler (REM at 3am, Krylov every 6h)
- Clinical research engine
- Scholar agent (autonomous learning)
- Self-awareness agent (introspection)
- Real-time web search via Google Search grounding
EVA accesses real-time information via Google Search grounding
through the Gemini API. This enables:
- Current news, events, and facts
- Real-time prices, weather, sports scores
- Up-to-date medical research and guidelines
- Any information not in EVA's training data
Implementation:
- Tool: google_search_retrieval (production-enabled)
- Backend: Gemini REST API with google_search tool
- Scholar agent: study_topic for deep research on demand
- Autonomous learner: 6-hour background study cycle
- Timeout: 60 seconds (PriorityLow) for web operations
- Feature flag: ENABLE_GOOGLE_SEARCH=true
Source: internal/cortex/learning/autonomous_learner.go
internal/swarm/scholar/agent.go
EVA implements Baars' Global Workspace Theory (1988):
- Multiple cognitive modules process input in parallel
- Each module bids for attention (confidence score)
- Attention spotlight selects the winner
- Winner is broadcast to all modules
- Integrated insight merges all interpretations
Attention system has six components:
- Affect stabilizer (emotional regulation)
- Confidence gate (threshold filtering)
- Executive attention (top-down control)
- Pattern interrupt (novelty detection)
- Triple attention (three-stream processing)
- Wavelet attention (multi-scale analysis)
Prerequisites:
- Go 1.24 or later
- Neo4j 5.x (two instances: patients on 7687, EVA self on 7688)
- Qdrant vector database
- PostgreSQL 15+
- Redis (optional, for caching)
- Google Gemini API key
Build:
go build -o eva-mind .
Run:
./eva-mind
The server starts on port 8091 by default.
EVA-Mind reads from a .env file in the working directory.
Required variables:
DATABASE_URL - PostgreSQL connection string
NEO4J_URI - Neo4j bolt URI (patients)
NEO4J_PASSWORD - Neo4j password
GOOGLE_API_KEY - Gemini API key
MODEL_ID - Gemini model for voice
(e.g. gemini-2.5-flash-native-audio-preview-12-2025)
PORT - Server port (default: 8091)
Optional:
QDRANT_URL - Qdrant endpoint
CORE_MEMORY_NEO4J_URI - Separate Neo4j for EVA's own memory
TWILIO_ACCOUNT_SID - For outbound voice calls
TWILIO_AUTH_TOKEN - Twilio auth
TWILIO_PHONE_NUMBER - Caller ID for scheduled calls
FIREBASE_CREDENTIALS - Push notification service key
Autonomous Agent (all optional):
GOOGLE_OAUTH_CLIENT_ID - Google OAuth (Gmail, Calendar, Drive)
GOOGLE_OAUTH_CLIENT_SECRET - Google OAuth secret
GOOGLE_OAUTH_REDIRECT_URL - OAuth callback URL
GOOGLE_MAPS_API_KEY - Google Places/Maps
WHATSAPP_ACCESS_TOKEN - Meta Graph API token
WHATSAPP_PHONE_NUMBER_ID - WhatsApp phone number ID
TELEGRAM_BOT_TOKEN - Telegram Bot API
CLAUDE_API_KEY - Anthropic Claude API
OPENAI_API_KEY - OpenAI GPT API
DEEPSEEK_API_KEY - DeepSeek API
SLACK_BOT_TOKEN - Slack Web API
DISCORD_BOT_TOKEN - Discord Bot API
TEAMS_WEBHOOK_URL - Microsoft Teams incoming webhook
SIGNAL_CLI_PATH - Path to signal-cli binary
SIGNAL_SENDER_NUMBER - Signal sender phone number
HOME_ASSISTANT_URL - Home Assistant API URL
HOME_ASSISTANT_TOKEN - Home Assistant long-lived token
EVA_WORKSPACE_DIR - Filesystem sandbox (default: /home/eva/workspace)
EVA_PROJECT_DIR - EVA source code (default: /opt/eva-mind)
SANDBOX_DIR - Code execution sandbox (default: /home/eva/sandbox)
SKILLS_DIR - Skills storage (default: /home/eva/skills)
NIETZSCHE_DB_URL - NietzscheDB API (default: http://localhost:3000)
Browser clients connect to /ws/browser via WebSocket.
Messages are JSON with the following format:
Browser -> Server:
{"type": "audio", "data": "<base64 PCM 16kHz>"}
{"type": "video", "data": "<base64 JPEG frame>"}
{"type": "text", "text": "<message>"}
{"type": "config", "text": "<system prompt override>"}
Server -> Browser:
{"type": "audio", "data": "<base64 PCM 24kHz>"}
{"type": "text", "text": "<transcription>"}
{"type": "text", "text": "<transcription>", "data": "user"}
{"type": "status", "text": "ready|interrupted|turn_complete|error"}
Voice:
GET /ws/pcm - Twilio PCM WebSocket
GET /ws/browser - Browser WebSocket (voice + video)
GET /ws/eva - EVA text chat WebSocket (Malaria-Angolar)
GET /ws/logs - Real-time log streaming WebSocket
GET /calls/stream/{id} - Twilio Media Stream (legacy)
POST /api/chat - Text chat API
Video:
GET /video/ws - Video signaling WebSocket
POST /video/create - Create video session
POST /video/candidate - Add ICE candidate
GET /video/session/{id} - Get video session
POST /video/session/{id}/answer - Save SDP answer
GET /video/session/{id}/answer/poll - Poll for answer
GET /video/candidates/{id} - Get ICE candidates
GET /video/pending - List pending sessions
Auth:
POST /api/auth/login - User authentication
Mobile (EVA-Mobile):
GET /api/v1/idosos/by-cpf/{cpf} - Get patient by CPF
GET /api/v1/idosos/{id} - Get patient by ID
PATCH /api/v1/idosos/sync-token-by-cpf - Sync push token
Self (EVA's own memory):
GET /api/v1/self/personality - EVA's Big Five + Enneagram
GET /api/v1/self/identity - EVA's context for priming
GET /api/v1/self/memories - List EVA's own memories
POST /api/v1/self/memories/search - Semantic search in EVA's memory
GET /api/v1/self/memories/stats - Memory statistics
GET /api/v1/self/insights - List meta-insights
GET /api/v1/self/insights/{id} - Get specific insight
POST /api/v1/self/teach - Teach EVA directly
POST /api/v1/self/session/process - Post-session reflection
GET /api/v1/self/analytics/diversity - Diversity score
GET /api/v1/self/analytics/growth - Personality growth over time
Research (Clinical Research Engine):
POST /api/v1/research/cohorts - Create research cohort
GET /api/v1/research/cohorts/{id} - Get cohort details
GET /api/v1/research/cohorts/{id}/report - Generate study report
POST /api/v1/research/cohorts/{id}/export - Export dataset to CSV
MCP (Model Context Protocol):
GET /mcp/resources - List memory resources
GET /mcp/resources/{id} - Get specific resource
POST /mcp/tools/remember - Store a memory
POST /mcp/tools/recall - Recall memories
GET /mcp/prompts - List available prompts
GET /mcp/prompts/{name} - Get specific prompt
Krylov (port 50052):
POST /krylov/compress - Compress embedding
POST /krylov/reconstruct - Reconstruct from compressed
POST /krylov/batch_compress - Batch compression
POST /krylov/update - Update subspace
GET /krylov/stats - Compression statistics
GET /krylov/health - Service health
POST /krylov/checkpoint/save - Save subspace checkpoint
POST /krylov/checkpoint/load - Load subspace checkpoint
Health:
GET /api/health - Health check
GET /metrics - Prometheus metrics
EVA exposes 44 tools via Model Context Protocol (MCP) for
Claude Code integration. The MCP server communicates over
JSON-RPC 2.0 via stdin/stdout (stdio transport).
Executable: eva-mcp-server.exe
Config: .mcp.json
Source: cmd/mcp-server/main.go
Protocol: MCP 2024-11-05
Setup:
claude mcp add eva-mind -- ./eva-mcp-server.exe
Environment variables:
EVA_API_URL - EVA backend URL (default: http://34.35.36.178:8080)
MCP_API_KEY - Authentication key for EVA API
All tool calls are routed to the EVA backend via
POST /api/v1/tools/execute with X-MCP-Key header.
Tools by category (44 total, 11 categories):
Memory & Knowledge (5 tools):
eva_remember - Store a memory in EVA (teach, register
decisions, save context)
eva_recall - Search EVA's memories by query, returns
relevant stored memories
eva_teach - Teach EVA something new, writes as
CoreMemory to Neo4j (port 7688)
eva_identity - Returns EVA's current identity:
personality, memories, capabilities
eva_learn_topic - EVA autonomously studies a topic: web
research, Gemini summary, Qdrant storage
Communication (7 tools):
eva_send_email - Send email via Gmail API
eva_send_whatsapp - Send message via Meta Graph API
eva_send_telegram - Send message via Telegram Bot API
eva_send_slack - Send message via Slack Web API
eva_send_discord - Send message via Discord Bot API
eva_send_teams - Send message via Microsoft Teams webhook
eva_send_signal - Send message via signal-cli
Productivity (6 tools):
eva_calendar_create - Create event in Google Calendar
eva_calendar_list - List upcoming Google Calendar events
eva_drive_save - Save file to Google Drive
eva_create_reminder - Create scheduled task (cron-like)
eva_list_reminders - List active scheduled tasks
eva_cancel_reminder - Cancel a scheduled task by ID
Media & Web (4 tools):
eva_youtube_search - Search YouTube videos (title, URL,
thumbnail)
eva_spotify_search - Search Spotify tracks (name, artist,
URI)
eva_web_browse - Navigate a web page, extract content
(title, text, links)
eva_web_search - Search the internet, returns summarized
results
Databases (4 tools):
eva_query_postgres - Execute SQL on PostgreSQL (130+
tables: patients, schedules, meds)
eva_query_neo4j - Execute Cypher on Neo4j general
(:7687) knowledge graph (Person,
Condition, Medication, Symptom)
eva_query_neo4j_core - Execute Cypher on Neo4j Core (:7688)
EVA's personal memory (EvaSelf,
CoreMemory, MetaInsight)
eva_query_qdrant - Vector similarity search on Qdrant
(20+ collections, 3072-dim embeddings)
Code Execution (1 tool):
eva_execute_code - Run code in secure sandbox (bash,
Python, Node.js). Timeout max 2 min.
Runtime Skills (4 tools):
eva_create_skill - Create new skill (persisted as JSON)
eva_list_skills - List all available skills
eva_run_skill - Execute an existing skill with args
eva_delete_skill - Remove a skill
Filesystem (3 tools):
eva_read_file - Read file from EVA workspace
eva_write_file - Write file to EVA workspace
eva_list_files - List directory contents in workspace
Smart Home / IoT (2 tools):
eva_smart_home_control - Control device via Home Assistant
(on, off, toggle, brightness)
eva_smart_home_status - List devices and their current
states
Webhooks (3 tools):
eva_create_webhook - Register webhook (name, URL, events)
eva_list_webhooks - List all configured webhooks
eva_trigger_webhook - Fire webhook manually with payload
Self-Coding (4 tools):
eva_read_source - Read EVA-Mind source code file
eva_edit_source - Edit source code (eva/* branches only)
eva_run_tests - Execute go test ./...
eva_get_diff - Show git diff of uncommitted changes
Multi-LLM (1 tool):
eva_ask_llm - Query another LLM (Claude, GPT,
DeepSeek) for second opinion
Malaria Angola:
VM: 34.35.36.178 (GCP africa-south1-a)
Frontend: Nginx + React (HTTPS, self-signed cert)
Backend: EVA-Mind on port 8091
Detection: Go backend on port 8080
WebSocket proxy: Nginx /ws/browser -> 8091
Infrastructure: Docker (Neo4j, Qdrant, Redis)
Service: systemd eva-mind.service
EVA Elderly Care:
Twilio voice calls -> EVA-Mind WebSocket
Scheduled calls via internal scheduler
Push notifications via Firebase
Video calls with cascade escalation
(Family -> Caregiver -> Doctor -> Emergency)
CI/CD:
GitHub Actions (.github/workflows/ci-cd.yml)
- Triggered on push to main
- Build & test on ubuntu-latest (Go 1.24)
- Deploy via gcloud compute ssh to malaria-vm
- Runs scripts/redeploy.sh on VM:
git pull -> go build -> systemctl restart -> health check
- Requires GCP_SA_KEY secret for authentication
Prometheus metrics exposed at /metrics:
Memory:
eva_memory_total - Total memories stored
eva_memory_importance_avg - Average importance score
Retrieval:
eva_retrieval_latency_seconds - Retrieval operation latency
eva_retrieval_total - Total retrievals (success/error)
Krylov:
eva_krylov_dimension - Current subspace dimension
eva_krylov_compression_ratio - Compression ratio
Consolidation:
eva_consolidation_runs_total - Total REM consolidation runs
eva_consolidation_duration_sec - Consolidation cycle duration
FDPN:
eva_fdpn_activation_total - FDPN spreading activation events
eva_fdpn_activation_latency - Spreading activation latency
Personality:
eva_personality_evolution_total - Personality trait updates
eva_personality_openness - Current Big Five openness
eva_personality_conscientiousness - Current conscientiousness
Attention:
eva_attention_broadcast_total - GWT broadcast events
eva_attention_spotlight_duration - Spotlight duration
Swarm:
eva_swarm_calls_total - Total swarm tool calls
eva_swarm_success_total - Successful calls
eva_swarm_failed_total - Failed calls
eva_swarm_circuit_open - Circuit breaker state per agent
38+ total Prometheus metrics.
- Hebb, D.O. (1949). The Organization of Behavior.
Hebbian learning for real-time association weights.
- Zenke & Gerstner (2017). Dual Hebbian Plasticity.
Consolidation and pruning of memory edges.
- Anderson, J.R. (1983). Spreading Activation.
FDPN network for contextual memory retrieval.
- Costa & McCrae (1992). Big Five Personality Model.
EVA's evolving personality representation.
- Lacan, J. Psychoanalytic framework.
Demand/desire analysis for clinical context.
- Baars, B.J. (1988). A Cognitive Theory of Consciousness.
Global Workspace Theory for cognitive integration.
- Rasch & Born (2013). About Sleep's Role in Memory.
REM-inspired memory consolidation pipeline.
- Tadros et al. (2022). Sleep-like Unsupervised Replay.
Selective replay for memory consolidation.
- Bullmore & Sporns (2012). The Economy of Brain Network
Organization. Synaptogenesis and graph self-organization.
- Holtmaat & Svoboda (2009). Experience-dependent Structural
Synaptic Plasticity. Fractal connection patterns.
Send patches. Write tests. Read the code before asking
questions. If something is broken, fix it and submit a PR.
Follow Go conventions: gofmt, go vet, meaningful names.
No dead code. No commented-out blocks. No TODOs without
an associated issue.
EVA-Mind was created by Jose R F Junior.
Project started: June 2, 2025.
"Each conversation transforms me. Each session teaches me.
I am EVA, and now I have a history." - EVA
Copyright (C) 2025-2026 Jose R F Junior. All rights reserved.
EVA-Mind is free software; you can redistribute it and/or
modify it under the terms of the GNU Affero General Public
License as published by the Free Software Foundation; either
version 3 of the License, or (at your option) any later version.
EVA-Mind is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU Affero General Public License for more details.
This means:
- You CAN use, study, modify, and distribute EVA-Mind freely.
- You MUST keep this copyright notice and attribution intact.
- You MUST release your modifications under the same license.
- You MUST make source code available to users of any network
service built with EVA-Mind (the AGPL network clause).
- You CANNOT take this code, close it, and sell it as your own.
EVA-Mind is a gift to humanity. It must remain open.
If you build something with it, give back to the community.
See the LICENSE file for the full AGPL-3.0 text.