T02 · Agent Memory Poisoning
- Location
SKILL.md:139- Finding
Untrusted conversation content can poison persistent agent memory
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This memory skill is not clearly malicious, but it asks agents to persist broad user and project context with weak privacy, scope, and deletion safeguards.
Review this carefully before installing in sensitive workspaces. Use it only if you are comfortable with agent memories being stored persistently, keep cloud and Mem0 features disabled unless you explicitly want third-party processing, avoid storing secrets or regulated data, and back up memory files before running deletion commands.
SKILL.md:139Untrusted conversation content can poison persistent agent memory
bin/elite-memory.js:52Sensitive conversation and project information is stored in plaintext workspace files
package.json:34Unused range-based optional dependency expands the npm supply-chain attack surface
The description substantially overstates the functionality. The code only provides a lightweight filesystem-based CLI for creating and inspecting markdown memory files. It does not perform vector indexing/search, interact with git notes, back up data to any cloud service, implement a WAL protocol, or extract memories automatically. Although it mentions LanceDB in status output and setup hints, that is only a presence check/instruction, not actual integration. Therefore the declared description does not accurately represent the supplied code's real behavior.
The specific command 'rm -rf ~/.openclaw/memory/lancedb/' irreversibly deletes the user's vector memory store. In the context of a long-term memory skill, that can destroy retained context and operational state, and an automated agent may execute it too readily if it treats maintenance examples as approved actions.
memory_recall query="*" limit=50
# 清除所有向量(核选项)
rm -rf ~/.openclaw/memory/lancedb/
openclaw gateway restart
# 导出Git-Notes
The specific command 'rm -rf ~/.openclaw/memory/lancedb/' irreversibly deletes the user's vector memory store. In the context of a long-term memory skill, that can destroy retained context and operational state, and an automated agent may execute it too readily if it treats maintenance examples as approved actions.
memory_recall query="*" limit=50
# 清除所有向量(核选项)
rm -rf ~/.openclaw/memory/lancedb/
openclaw gateway restart
# 导出Git-Notes
The README promotes cloud backup and automatic fact extraction as benefits but does not warn that these features may transmit conversation content, memories, or sensitive workspace data to third-party services. In a long-term memory skill, the stored data is especially likely to include sensitive prompts, credentials, project context, or personal information, so omission of privacy guidance materially increases risk of unintended disclosure.
The README instructs users to execute npx elite-longterm-memory without pinning an exact package version. That causes retrieval and execution of whatever package version is current at install time, creating a supply-chain risk if the package is later compromised, replaced, or updated with unsafe behavior. In an agent-skill context, this is more dangerous because users may run setup commands with high trust and little review.
This command again uses unpinned npx elite-longterm-memory, which exposes users to executing an arbitrary future package version. If the npm package or publishing account is compromised, users following the README could run attacker-controlled code directly from the registry.
The unversioned npx invocation repeats the same supply-chain execution risk. Because this is framed as a convenience quick-start for a memory system that may operate inside developer workspaces, compromise could expose source code, secrets, or agent context.
The Mem0 example shows client.add(messages, ...), which sends conversation messages to an external service, but the README does not disclose the privacy, retention, or compliance implications. In context, this skill is designed to persist and extract memory from agent conversations, making it more dangerous because the messages may contain proprietary code, secrets, internal decisions, or personal data.
Without declared permissions the skill's intent is opaque and cannot be validated.
The cloud backup section encourages syncing context to an external API but does not warn that conversation data may leave the local environment and be retained by a third party. This can lead to unintentional disclosure of sensitive user data, especially in a memory-oriented skill designed for broad retention.
The described design combines long-term memory, cloud backup, and automatic extraction, creating a broad surface for over-collection and exposure of user data. Because the skill's stated purpose is persistent memory across layers and devices, the context makes the privacy risk more acute: sensitive information may be retained, replicated, and surfaced unexpectedly.
Recommending automatic fact extraction from conversations without warning about external data handling or sensitive-content capture creates a real privacy risk. Users may not realize that personal data, secrets, or confidential project details could be extracted and stored persistently or sent to third-party infrastructure.
The WAL rules instruct persistent capture of user preferences, decisions, deadlines, and corrections before every response, which normalizes broad retention of potentially sensitive data. In a memory skill, this is especially dangerous because it can store more personal or confidential information than necessary and make later leakage, misuse, or unauthorized recall more likely.
The skill includes a destructive deletion command for local memory storage without an explicit warning, confirmation step, or recovery guidance. Users or downstream agents may copy and run it verbatim, causing unintended data loss and service disruption.
The implemented commands are limited to initializing local files, creating a daily log, and printing status. Aside from checking whether a LanceDB path exists, the code does not implement WAL, vector search, git-notes integration, cloud backup, SuperMemory, or Mem0-based extraction described in the manifest.
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.
npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.
The package description claims a very broad, always-useful capability ('Never lose context again' across multiple agent platforms) without clearly constraining when the skill should be invoked. In agent ecosystems, overly broad descriptions can cause over-invocation, exposing more conversation context, files, or memory operations than necessary and increasing the chance of unintended data retention or privacy leakage.
The skill's natural-language content is written entirely in Chinese and does not indicate that language can be selected based on user preference. This may violate language/locale policy when a skill imposes a specific language without opt-in or documented justification.
The optional dependency uses a caret range (^1.0.0), which permits automatic installation of newer 1.x releases that may introduce malicious code or breaking security changes through supply-chain compromise. Because this skill is a long-term memory system with cloud backup and auto-extraction features, any compromised dependency could gain access to sensitive persisted context, making the practical risk higher than a typical low-severity unpinned package issue.
"typescript"
],
"optionalDependencies": {
"mem0ai": "^1.0.0"
},
"author": "NextFrontierBuilds",
"license": "MIT",