Tools
NeuralNexus
By Advanced memory & preference plugin for OpenClaw. Features category-aware exponential decay (facts, entities, decisions) and semantic preference overwriting via Qdrant & Xenova CPU embeddings. Prevents duplicates with a 0.92 similarity threshold. Intelligent auto-capture ensures your agent learns and forgets with context-aware persistence.
Install
npm install
npmConfiguration Example
{
"memory": {
"slot": "neural_nexus"
}
}README
<p align="center">
<img src="./images/NeuralNexus_txt_logo.gif" alt="Neural Nexus Logo" width="320" />
</p>
<h1 align="center">Neural Nexus Plugin for OpenClaw</h1>
<p align="center">
<a href="#"><img alt="Version" src="https://img.shields.io/badge/version-2026.2.17-0ea5e9"></a>
<a href="#"><img alt="TypeScript" src="https://img.shields.io/badge/TypeScript-5.9+-3178C6?logo=typescript&logoColor=white"></a>
<a href="#"><img alt="OpenClaw" src="https://img.shields.io/badge/OpenClaw-Memory_Plugin-10b981"></a>
<a href="#"><img alt="Vector DB" src="https://img.shields.io/badge/Qdrant-Vector_DB-ff4d4f"></a>
<a href="#"><img alt="SQLite Audit" src="https://img.shields.io/badge/SQLite-Replacement_Audit-2563eb"></a>
</p>
Neural Nexus is a long-term memory plugin for OpenClaw. It helps an AI remember useful information across conversations by storing and retrieving meaning-based memories.
## What Neural Nexus Is
Neural Nexus is a long-term memory plugin for OpenClaw. It helps an AI remember useful information across conversations by storing and retrieving meaning-based memories.
## 1) Semantic Memory Storage
Neural Nexus converts text into vectors (numeric representations of meaning) and stores them in Qdrant.
Why this matters:
- It can find related ideas even when wording is different.
- Retrieval is smarter than plain keyword matching.
## 2) Automatic Memory Capture
After successful agent interactions, the plugin can automatically decide what to save.
Key points:
- Controlled by `autoCapture` (default: on).
- Runs on `agent_end` events.
- Ignores failed or invalid events.
## 3) Built-in Recall Tool (`memory_recall`)
Neural Nexus registers a tool the AI can call to search long-term memory.
Inputs:
- `query` (required)
- `limit` (optional, 1-20, default 5)
Output:
- Matching memories with adjusted scores.
## 4) Memory Categories
Each memory is categorized as:
- `preference`
- `fact`
- `entity`
- `decision`
- `other`
Why this matters:
- Different categories are handled differently for decay and reinforcement.
## 5) Category-Based Decay
Memories fade at different rates depending on category (`lambda`).
Current behavior:
- `preference`: `lambda = 0` (no decay)
- `fact`: slow decay
- `entity`: medium-slow decay
- `decision`: faster decay
- `other`: fastest decay
## 6) Time-Adjusted Memory Ranking
Recall scores are adjusted using:
- similarity score
- last access time
- decay rate (`lambda`)
- memory strength
Result:
- More relevant and recently reinforced memories rank higher.
## 7) Preference Overwrite Logic
Preferences are treated as stable memory.
Behavior:
- If a new preference is very similar to an existing one (high threshold around `0.92`), it overwrites the old memory instead of adding a duplicate.
- Stored preferences have `lambda = 0` and do not decay.
## 8) Duplicate Prevention for Non-Preferences
For non-preference memories, Neural Nexus avoids storing near-duplicates.
Result:
- Cleaner memory set with less repetition.
## 9) Reinforcement on Recall
When a memory is recalled:
- `last_accessed` is updated.
- For reinforce-enabled categories, `strength` increases (`+0.05`).
- If `lambda` exists, it is slightly reduced (`* 0.98`), making future decay slower.
Note:
- Preferences are excluded from reinforcement because they are already non-decaying.
## 10) Conversation Consolidation
For longer conversations, the plugin can combine recent user messages into one memory candidate.
Controls:
- `consolidation` (default: on)
- `consolidationThreshold` (default: 4 user messages)
Benefits:
- Captures meaningful context without storing too many small fragments.
## 11) Candidate Quality Filtering
Neural Nexus filters memory candidates to reduce noise.
It does things like:
- Focus on user-role messages.
- Ignore very short/low-value content.
- Normalize whitespace.
- Limit candidate length (~1200 chars).
## 12) Rule-Based Category Detection
Category detection uses simple language cues.
Examples:
- "prefer/like/dislike" -> `preference`
- "decided/chose/instead of" -> `decision`
- "is a/called" -> `entity`
- Otherwise -> `fact`
## 13) Initialization Safety Checks
Before capture/recall, the plugin verifies that:
- Embedding model is initialized.
- Qdrant collection exists and matches vector dimensions.
If setup fails, it logs errors and avoids unsafe operations.
## 14) Config Normalization + Defaults
If config values are missing, Neural Nexus applies defaults.
Important defaults:
- Model: `Xenova/bge-small-en-v1.5`
- Device: `cuda` (or `cpu` if configured)
- Qdrant URL: `http://localhost:6333`
- Collection: `openclaw_memories`
- Auto-capture/recall/consolidation: enabled
- Consolidation threshold: `4`
Also supports `${ENV_VAR}` substitution in config strings.
## 15) E5 Model Compatibility
If the embedding model name includes `e5`, text is prefixed with `query: ` to match E5 retrieval conventions.
Why this matters:
- Better search quality for E5-style models.
## 16) Service Registration
The plugin registers itself as memory service ID `neural_nexus` and logs when active.
## 17) Automatic Qdrant Collection Setup
At startup, it checks whether the target collection exists.
If not, it creates one with cosine distance and the correct vector size.
## 18) Memory Replacement Audit (SQLite)
When a memory is replaced with a newer memory (for example, preference overwrite), Neural Nexus can record that replacement in SQLite.
What gets logged:
- memory id
- category
- old text
- new text
- similarity score
- replacement timestamp
Default path:
- `./data/neural_nexus_replacements.sqlite`
## Installation
1. Copy the plugin to your OpenClaw extensions directory: `~/.openclaw/extensions/neural_nexus`
2. Configure OpenClaw to use `neural_nexus` as your memory slot in `openclaw.json`:
```json
{
"memory": {
"slot": "neural_nexus"
}
}
```
## Configuration
```json
{
"memory": {
"slot": "neural_nexus",
"pluginConfig": {
"embedding": {
"model": "Xenova/bge-small-en-v1.5",
"device": "cuda"
},
"qdrant": {
"url": "http://localhost:6333",
"collection": "openclaw_memories",
"apiKey": "${QDRANT_API_KEY}"
},
"autoCapture": true,
"autoRecall": true,
"consolidation": true,
"consolidationThreshold": 4,
"replacementLog": {
"enabled": true,
"sqlitePath": "./data/neural_nexus_replacements.sqlite"
}
}
}
}
```
## Requirements
- Node.js v18+
- A running Qdrant instance
## Development
```bash
npm install
npm run build
```
## Bottom Line
Neural Nexus gives an AI practical long-term memory:
- smart semantic search
- automatic capture and recall
- category-aware forgetting
- stable preference tracking
- duplicate control
- reinforcement for frequently used memories
tools
Comments
Sign in to leave a comment