nima-core（尼玛核心）

NIMA Core 2.3

Neural Integrated Memory Architecture — A complete memory system for AI agents with emotional intelligence.

Website: https://nima-core.ai GitHub: https://github.com/lilubot/nima-core

🚀 Quick Start

# Install
pip install nima-core

# Or with LadybugDB (recommended for production)
pip install nima-core[vector]

# Set embedding provider
export NIMA_EMBEDDER=voyage
export VOYAGE_API_KEY=your-key

# Install hooks
./install.sh --with-ladybug

# Restart OpenClaw
openclaw restart

🔒 Privacy & Permissions

Data Access:

• ✅ Reads session transcripts from ~/.openclaw/agents/*/sessions/*.jsonl
• ✅ Writes to local storage at ~/.nima/ (databases, affect history, embeddings)

Network Calls (conditional on embedder choice):

• 🌐 Voyage API — Only when NIMA_EMBEDDER=voyage (sends text for embeddings)
• 🌐 OpenAI API — Only when NIMA_EMBEDDER=openai (sends text for embeddings)
• 🔒 Local embeddings — Default (NIMA_EMBEDDER=local), no external API calls

Opt-in Controls:

// openclaw.json
{
  "plugins": {
    "entries": {
      "nima-memory": {
        "enabled": true,
        "skip_subagents": true,      // Exclude subagent sessions (default)
        "skip_heartbeats": true,      // Exclude heartbeat checks (default)
        "noise_filtering": {
          "filter_heartbeat_mechanics": true,
          "filter_system_noise": true
        }
      }
    }
  }
}

Privacy Defaults:

• Subagent sessions excluded
• Heartbeat/system noise filtered
• Local embeddings (no external calls)
• All data stored locally

To disable: Remove nima-memory from plugins.allow in openclaw.json

What's New in 2.1

VADER Affect Analyzer

• Contextual Analysis: Caps boost (1.5x), punctuation emphasis (!!!), negation handling, degree modifiers
• 30+ Idiom Recognition: Understands phrases like "not bad", "kind of", "sort of"
• Panksepp 7-Affect Mapping: Direct mapping from VADER sentiment to SEEKING, RAGE, FEAR, LUST, CARE, PANIC, PLAY
• Guardian Archetype Transformation: User anger → Agent concern/care response modulation
• Replaces previous lexicon-based emotion detection

Noise Remediation (4-Phase)

1. Empty Validation — Filters out null/empty messages
2. Heartbeat Filters — Excludes system noise (HEARTBEAT_OK, polling messages)
3. Deduplication — Removes duplicate content within sessions
4. Metrics Collection — Tracks capture quality and filter effectiveness

Performance Improvements

• LadybugDB Circular Import Fix: Resolved import issues in LadybugDB backend
• Increased Token Budget: Recall budget increased from 500 to 3000 tokens
• Connection Pooling: Improved connection management for LadybugDB backend

What's New in 2.0

LadybugDB Backend

• 3.4x faster text search (9ms vs 31ms)
• Native vector search with HNSW (18ms)
• 44% smaller database (50MB vs 91MB)
• Graph traversal with Cypher queries

Security Hardened

• Query sanitization (FTS5, SQL injection prevention)
• Path traversal protection
• Temp file cleanup
• Error handling throughout

Thread Safe

• Singleton pattern with double-checked locking
• API timeouts (30s Voyage, 10s LadybugDB)
• Connection pooling ready

348 Tests

• Full unit test coverage
• Thread safety verified
• Edge cases covered

Architecture

OPENCLAW HOOKS
├── nima-memory      — Three-layer capture with 4-phase noise remediation
├── nima-recall-live — Lazy recall injection (before_agent_start)
└── nima-affect      — VADER-based real-time affect analysis

PYTHON CORE
├── nima_core/cognition/
│   ├── dynamic_affect.py       — Panksepp 7-affect system
│   ├── personality_profiles.py — JSON personality configs
│   ├── vader_affect.py         — VADER sentiment analyzer (NEW v2.1)
│   └── archetypes.py           — Baseline affect profiles
└── scripts/
    ├── nima_ladybug_backend.py — LadybugDB CLI
    └── ladybug_parallel.py     — Parallel migration

DATABASE (SQLite or LadybugDB)
├── memory_nodes   — Messages with embeddings
├── memory_edges   — Graph relationships
└── memory_turns   — Conversation turns

Performance

API

from nima_core import DynamicAffectSystem, get_affect_system
from nima_core.cognition.vader_affect import VaderAffectAnalyzer

# Get singleton instance (thread-safe)
affect = get_affect_system(identity_name="lilu")

# Process input and get affect state
state = affect.process_input("I'm so excited about this project!")
print(state.current)  # {"SEEKING": 0.72, "PLAY": 0.65, ...}

# Use VADER analyzer directly
analyzer = VaderAffectAnalyzer()
result = analyzer.analyze("This is AMAZING!!!")
print(result.affects)  # {'PLAY': 0.78, 'SEEKING': 0.71, ...}

# Recall memories (via hooks - automatic)
# Or manually via CLI:
# nima-query who_search "David" --limit 5
# nima-query text_search "project" --limit 5

Configuration

Hooks

nima-memory (Capture)

• Captures input, contemplation, output on every turn
• 4-phase noise remediation (empty validation, heartbeat filters, dedup, metrics)
• Stores to SQLite or LadybugDB
• Computes and stores embeddings

nima-recall-live (Recall)

• Injects relevant memories before agent starts
• Lazy loading — only top N results
• Deduplicates with injected context
• Token budget: 3000 (increased from 500 in v2.1)

nima-affect (Emotion)

• VADER-based real-time affect analysis from text
• Contextual analysis (caps, punctuation, negation, degree modifiers)
• 30+ idiom recognition
• Maintains Panksepp 7-affect state
• Guardian archetype transformation (user anger → agent care)

Installation Options

SQLite (Development)

pip install nima-core
./install.sh

LadybugDB (Production)

pip install nima-core[vector]
./install.sh --with-ladybug

Documentation

Security & Privacy

Data Access

This plugin accesses:

• ~/.openclaw/agents/.../*.jsonl — Session transcripts (for memory capture)
• ~/.nima/ — Local memory database (SQLite or LadybugDB)
• ~/.openclaw/extensions/ — Hook installation

Network Calls

Embeddings are sent to external APIs:

• Voyage AI (api.voyageai.com) — Default embedding provider
• OpenAI (api.openai.com) — Optional embedding provider
• Local — No external calls when using sentence-transformers

Required Environment Variables

Installation Script

The install.sh script:

1. Checks for Python 3 and Node.js
2. Creates ~/.nima/ directories
3. Installs Python packages via pip
4. Copies hooks to ~/.openclaw/extensions/

No external downloads. All packages come from PyPI.

Changelog

v2.1.0 — VADER Affect Analyzer (Feb 17, 2026)

• Added: VADER-based affect analyzer replacing lexicon-based detection

- Contextual analysis: caps boost (1.5x), punctuation (!!!), negation, degree modifiers - 30+ idiom recognition - Panksepp 7-affect mapping (SEEKING, RAGE, FEAR, LUST, CARE, PANIC, PLAY) - Guardian archetype transformation (user anger → agent concern/care)

• Added: 4-phase noise remediation (empty validation, heartbeat filters, dedup, metrics)
• Fixed: LadybugDB circular import issue
• Changed: Recall token budget increased from 500 to 3000
• Improved: Connection pooling for LadybugDB backend

v2.0.3 — Security Hardening (Feb 15, 2026)

• Security: Fixed path traversal vulnerability in affect_history.py (CRITICAL)
• Security: Fixed temp file resource leaks in 3 files (HIGH)
• Fixed: Corrected non-existent json.JSONEncodeError → TypeError/ValueError
• Improved: Exception handling - replaced 5 generic catches with specific types
• Quality: Better error visibility and debugging throughout

v2.0.1 — Thread Safety + Metadata

• Fixed: Thread-safe singleton with double-checked locking
• Security: Clarified metadata requirements (Node.js, env vars)
• Docs: Added security disclosure for API key usage

v2.0.0 — LadybugDB + Security

• Added: LadybugDB backend with HNSW vector search
• Added: Native graph traversal with Cypher
• Added: nima-query CLI for unified queries
• Security: SQL/FTS5 injection prevention
• Security: Path traversal protection
• Security: Temp file cleanup
• Fixed: Thread-safe singleton initialization
• Fixed: API timeouts (Voyage 30s, LadybugDB 10s)
• Tests: 348 tests passing
• Performance: 3.4x faster text search, 44% smaller DB

v1.2.1 — Consciousness Architecture

• Added: 8 consciousness systems (Φ, Global Workspace, self-awareness)
• Added: Sparse Block VSA memory
• Added: ConsciousnessCore unified interface

v1.1.9 — Hook Efficiency Fix

• Fixed: nima-recall hook spawning new Python process every bootstrap
• Performance: ~50-250x faster hook recall

v1.2.0 — Affective Response Engines

• Added: 4 Layer-2 composite affect engines
• Added: Async affective processing
• Added: Voyage AI embedding support