Memory Security
for AI Agents
ShieldCortex helps AI agents remember useful things, stop dangerous memory from spreading, and show operators exactly what was captured, recalled, blocked, and reviewed. Start free, get a 14-day Pro trial on first install, and only add cloud when you need team visibility.
Available on npm and PyPI
ShieldCortex v4.4.0
X-Ray — inspect packages, files, and plugins for hidden risk. CI/CD gating, file watch, pre-install hooks, and memory guard. Plus Dream Mode, Hybrid Recall, Dependency Scanner, and more.
See It In Action
Watch ShieldCortex block a prompt injection and privilege escalation in real time.
Also on X/Twitter
Stop Bad Memory
Catch poisoning before it becomes future truth.
Prompt injection, hostile instructions, fragmented payloads, suspicious provenance, and leaked credentials are scanned before they land in durable agent memory.
Inspect Recall
See what the agent stored and why it ranks.
Capture, Recall, and Review turn memory into an operator workflow. Inspect stored memories, rank explanations, contradictions, duplicates, and low-trust records before they shape output.
Operator Security
Keep a human in control when agents misbehave.
Review queues, Incident Replay, Device Doctor, Verify, and Iron Dome analytics show what happened, what was blocked, what synced, and what the operator should do next.
Iron Dome New
The defence pipeline protects your agent's memory. Iron Dome protects your agent's actions.
Injection Scanner
40+ patterns across 8 attack categories
Action Gate
Approve, require approval, or block actions
PII Guard
Block protected data from agent output
Instruction Gateway
Only trusted channels can issue commands
Kill Switch
Trigger phrase pauses memory creation instantly
Sub-Agent Control
Restrict spawned agents' capabilities
Every memory your OpenClaw agent saves passes through ShieldCortex. Prompt injection, credential leaks, poisoned context — caught before it reaches storage. New: the Tool Call Interceptor actively blocks suspicious tool calls and requires operator approval before proceeding. Works as a standalone memory system or as a security layer for any memory backend.
Hook Integration
Hook and plugin install separately. Auto-memory is opt-in with smart deduplication.
Real-time Plugin
NewScans every LLM input for threats and auto-extracts memories from outputs. Fire-and-forget — never blocks your agent.
Install directly with openclaw plugins install @drakon-systems/shieldcortex-realtime or use the ShieldCortex wrapper on existing setups.
Requires OpenClaw v2026.2.15+. Details →
Works on macOS, Linux (including headless servers), and Windows. Rolling out to a paid team? Use the install commands, activate the Team licence locally, set the Cloud API key, enable cloud sync, and run the persistent worker on always-on boxes. No separate signup is needed on the target machine.
Team setup guide →AI Agents Have Three Security Problems
Most tools add memory. Very few make that memory trustworthy.
Amnesia
Your agent forgets everything between sessions. No context. No learning. Starting from zero every time.
Memory Vulnerability
The moment you add memory, you create an attack surface. Memory poisoning. Prompt injection. Credential harvesting.
Behaviour Vulnerability
Agents can be tricked into executing harmful actions via prompt injection. No guardrails on what they do, only what they remember.
ShieldCortex = Memory + Poisoning Defence + Operator Control
Not just memory for agents. Memory security for agents.
What You Get
Full protection for free. Pro licence for custom rules. Team licence for cloud sync.
Local Protection
npm package — free and open source
- Memory Firewall — Block injection attacks in real time
- Trust Scoring — Filter by source reliability
- Anomaly Detection — Behavioural scoring over time
- Codex MCP support — one install covers Codex CLI and the Codex VS Code extension on the same machine
- MCP Tools — Native support for Claude Code, OpenClaw
- Universal Memory Bridge — Guard any memory backend with the defence pipeline
- 6-Layer Defence Pipeline — Sanitise, detect, analyse, validate, score, credential scan
- Security Audit — A-F grading, CVE detection, CI/CD integration via GitHub Action
- Iron Dome — Behaviour protection: injection scanning, action gating, PII guard
- Memory Types — Typed memories: user, feedback, project, reference. Organise what your agent remembers.
- Staleness Scoring — Age-based decay with automatic warnings on old memories. Stale data flagged, not trusted blindly.
- LLM-Powered Reranking — Hybrid recall: embedding search + optional LLM reranker for precision retrieval.
- Dream Mode — Background memory consolidation: merge duplicates, archive stale, detect contradictions. Like brain sleep for agents.
- Save Filtering — Prevents saving derivable information (file paths, git refs, env vars). Only stores what matters.
- Docker Install Safety — Auto-detects containers and prevents broken installs that damage host systems.
Context Preservation
- Pre-Compaction Extraction — Auto-saves decisions, fixes, and learnings before context is lost
- Session Continuity — Auto-loads project context at session start, extracts on exit
- Salience Scoring — Prioritises what matters, lets noise decay naturally
- Memory Consolidation — Promotes short-term to long-term memory, like brain sleep
Pro + Team
Licence key — from £29/mo
Pro Features
- Custom Injection Patterns — Up to 50 regex rules for domain-specific threats
- Custom Iron Dome Policies — Tailor behaviour controls to your use case
- Audit Export — JSON/CSV export of your defence logs
- Skill Scanner Deep Mode — Multi-file semantic analysis
- Dependency Scanner — Detect malicious packages, typosquats, and suspicious install scripts in your project
Team Features
- Cloud Audit Sync — Centralised logs across all devices
- Team Management — Invite members, shared custom patterns
- Memory Scopes — Private vs team memories for multi-agent deployments. Control who sees what.
- Positive Feedback Capture — Cortex now captures confirmations, not just mistakes. Learn from success AND failure.
Your AI agent runs npm install.
Who's watching?
AI coding agents execute package installs autonomously. ShieldCortex catches supply chain attacks that Snyk and Dependabot miss.
Zero-Day Detection
- Catches attacks BEFORE they're in CVE databases
- Blocklist of known malicious packages updated in real-time
-
Found what others missed:
axios 1.14.1,plain-crypto-js,color-diff-napi
Typosquat & Script Analysis
- Flags lookalike package names (Levenshtein distance)
-
Scans
postinstallscripts for payload downloads, OS detection, credential access - Warns about brand-new packages with install hooks
Auto-Quarantine
- Doesn't just warn — actively removes threats
- Moves malicious packages to quarantine with full manifest
-
shieldcortex audit --deps --auto-protectfor CI/CD pipelines
Your Agent Never Forgets
AI agents lose context during compaction, between sessions, and when conversations get long. ShieldCortex intercepts at every point where memory is lost.
Before Compaction
Auto-extracts decisions, fixes, and learnings before they get compressed away.
Session Start
Loads your highest-priority memories automatically. No blank slate.
Smart Priority
Salience scoring keeps architecture decisions and bug fixes; lets noise decay naturally.
Consolidation
Promotes short-term to long-term memory and merges related items. Like brain sleep.
Get Started in Minutes
Choose your path.
OpenClaw Users
Install the session hook
Install the real-time plugin
Reload OpenClaw and start protected sessions
Python
NewLangChain/CrewAI extras available
Local scanning — no API key needed
Node.js
Install the ShieldCortex CLI
Configure your agent integration
View threats, audit log, and brain visualisation
Node.js 18+ or Python 3.9+ — macOS, Linux & Windows
ShieldCortex X-Ray
See what your AI can't. Inspect packages, plugins, and files for hidden instructions, exfiltration paths, covert payloads, and AI-targeted exploits.
Start Protecting Your Agents
Install free. Upgrade locally with a licence key.