NeuroCache Pro gives AI agents graph-based associative memory for discovering relationships, tracing causes, resolving contradictions, and reusing knowledge across projects.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install neurocache-pro
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install neurocache-pro using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is NeuroCache Pro?
NeuroCache Pro is a biological-inspired neural memory system for AI agents and Openclaw Skills. Instead of relying only on keyword or vector similarity, it represents memories as neurons connected by typed synapses, then uses spreading activation to discover conceptual, temporal, causal, semantic, emotional, and hierarchical relationships—even when two memories share no obvious wording or embedding overlap.
The skill combines Hebbian reinforcement, configurable decay, contradiction detection, depth-controlled recall, graph pruning, brain snapshots, rollback, and cross-project transplantation. Its core retrieval algorithms use graph traversal and other deterministic techniques without requiring an LLM API, while optional automatic extraction can use an LLM or regex-based patterns.
NeuroCache Pro Use Cases
- Trace multi-step causal chains, such as why a deployment failed or which configuration caused an incident.
- Discover cross-domain relationships that ordinary vector search may miss, such as links between authentication, caching, tokens, and invalidation.
- Preserve long-term project decisions, facts, preferences, errors, workflows, TODOs, and user instructions.
- Detect conflicts between old and new facts, prioritize current information, and retain outdated context as a historical note.
- Manage growing knowledge graphs with activation thresholds, decay, pruning, partitioning, and depth-based retrieval.
- Create safety checkpoints before risky memory operations, compare graph changes, and roll back to a prior snapshot.
- Transfer high-value decisions and lessons from one project brain to another while preserving relationships and provenance.
- Give Openclaw Skills and compatible AI agents durable, structured memory across sessions.
How NeuroCache Pro Works
- Initialize a local neural-memory brain backed by SQLite and expose its tools through MCP.
- At session start, load recent context with
nmem_contextand recall topic-specific memories withnmem_recall. - Store facts, decisions, errors, preferences, insights, workflows, and other durable information with
nmem_remember. - Convert memories into entities and typed synapses, including causal, temporal, semantic, emotional, hierarchical, reference, location, and action relationships.
- Activate related neurons through graph traversal and spreading activation rather than depending only on lexical or embedding similarity.
- Reinforce frequently co-accessed connections using Hebbian learning while applying configurable decay to stale, low-priority memories.
- Detect semantic conflicts through
CONTRADICTS, compare timestamps and priority values, downgrade likely-obsolete memories, and flag ambiguous conflicts for user confirmation. - Select a recall depth from 0 to 3: immediate facts, standard context, habits and patterns, or deep cross-domain and causal exploration.
- Automatically extract important facts, decisions, errors, and TODOs from conversation segments with
nmem_autoat the end of a session. - Monitor graph health, create snapshots, inspect diffs, roll back changes, and transplant filtered knowledge between project brains.
NeuroCache Pro Setup
Requirements
- Python 3.9 or newer
- Windows, macOS, or Linux
- An AI agent that supports SKILL.md instructions and MCP, such as Claude Code, Cursor, Codex, or Gemini CLI
- SQLite, included through the Python runtime
Install neural memory
pip install neural-memory
nmem init
This creates the default brain under ~/.neuralmemory/ and prepares the MCP integration.
Configure MCP
Add the following server configuration to the supported agent's MCP settings:
{
"mcpServers": {
"neural-memory": {
"command": "python3",
"args": ["-m", "neural_memory.mcp"],
"env": {
"NEURALMEMORY_BRAIN": "default"
}
}
}
}
Verify the installation
nmem stats
The output should report brain statistics, including neuron, synapse, and fiber counts. Use nmem_health for diagnostics if MCP connectivity or recall fails.
Core graph retrieval does not require an API key. An optional embedding provider can improve semantic matching, and nmem_auto may optionally use an LLM for automatic memory extraction; regex pattern matching is available as an alternative.
NeuroCache Pro Data Schema & Taxonomy
NeuroCache Pro organizes knowledge into a brain composed of entities and typed synapses.
Core records
| Record | Fields | Purpose |
|---|---|---|
| Entity | id, content, type, tags, priority, created, last_accessed, decay_score |
Stores a memory, fact, decision, event, preference, or other knowledge item. |
| Synapse | from_id, type, to_id, strength, created, last_fired |
Connects entities and records the strength and usage of their relationship. |
| Brain | Project or domain identifier | Separates knowledge stores and enables controlled transplantation. |
| Snapshot | Label, graph state, timestamp | Supports diffs and rollback before or after risky operations. |
Memory taxonomy
Supported memory types include fact, decision, preference, todo, insight, context, instruction, error, workflow, and reference. Priority ranges from 0 to 10; higher-priority memories decay more slowly, with a default priority of 5. TODO items expire after 30 days.
Synapse taxonomy
The system supports 20 typed relationships across these groups: BEFORE and AFTER for time; CAUSED_BY and LEADS_TO for causality; IS_A, HAS_PROPERTY, and PART_OF for semantics; FELT and EVOKES for emotion; CONTRADICTS for conflicts; SIMILAR_TO and RELATED_TO for general association; PARENT_OF and CHILD_OF for hierarchy; REFERENCES and CITED_BY for citations; LOCATED_AT and CO_OCCURRED for location and co-occurrence; and PERFORMED_BY and RESULTED_IN for actions.
Generated and managed data
- Local brain data is created under
~/.neuralmemory/by default. - SQLite stores the memory graph and its metadata.
- Contradiction links preserve competing claims instead of silently deleting data.
- Decay metadata includes activation thresholds, access timestamps, and reinforcement state.
- Transplanted records retain provenance such as
transplanted_from. - Version commands generate snapshots, diffs, and rollback states.
Recall depth
| Depth | Mode | Typical use |
|---|---|---|
| 0 | Immediate | Fast facts and recent context; under 10 ms target latency. |
| 1 | Context | Standard recall; approximately 50 ms target latency and the recommended default. |
| 2 | Habits | Pattern matching and workflow suggestions; approximately 200 ms target latency. |
| 3 | Deep | Cross-domain relationships and causal chains; approximately 500 ms target latency. |
NeuroCache Pro Advanced Features
- Spreading activation graph traversal finds conceptual associations without requiring keyword or embedding overlap.
- Hebbian learning reinforces connections between frequently co-accessed memories.
- Twenty typed synapse classes support causal reasoning, timelines, hierarchy, emotion, references, actions, and general association.
- Automatic
CONTRADICTSdetection compares timestamps and priorities, downgrades stale claims, and preserves uncertainty for confirmation. - Configurable decay supports
initialDecay,intervalDays,minActivation,pruneThreshold,reinforceFactor, automatic pruning, and weekly schedules. - Four recall depths balance speed, graph exploration, and token consumption; depth 1 is recommended for routine queries and depth 3 for complex reasoning.
- Graph expansion controls include pruning, partitioning, threshold tuning, and migration to SQLite-backed storage for larger brains.
nmem_autoextracts durable facts, decisions, errors, and TODOs from conversation text, with optional LLM assistance or regex-based extraction.nmem_habitsidentifies repeated work patterns and suggests workflows.nmem_versionsupports labeled snapshots, graph diffs, and one-command rollback.nmem_transplantfilters and migrates selected memory types, priorities, tags, and relationship structures between project brains.- MCP integration makes the memory tools available to compatible Openclaw Skills and other SKILL.md-driven agents.
- The core algorithm is deterministic and LLM-independent, reducing inference cost while preserving optional semantic and extraction enhancements.
SKILL.md
Loading
Related Openclaw Skills
Business Automation Hub
Business Automation Hub helps small teams decide what to automate, build reliable workflows, quantify ROI, and detect silent failures.
Cron Scheduler Pro
Cron Scheduler Pro turns natural-language scheduling intentions into reliable, auditable, and locally stored task execution contracts for AI Agents.
Ticket Revenue Settlement Check (Free)
A free local checker that recomputes ticket revenue, channel commissions, refunds, and settlement totals line by line.
Grocery Price Book
Grocery Price Book turns receipts into personalized price intelligence for comparing stores, detecting shrinkflation, and making smarter bulk-buy decisions.
AgentVibes Content Creation and Management Skill
AgentVibes Content Creation and Management connects an AI coding agent to content-generation APIs for automated multimodal creation, review, formatting, and publishing.
Agent Telegram LITE
Agent Telegram LITE standardizes how main, backend, and frontend AI agents send task updates to a fixed Telegram recipient.
Featured*
