Back to plugin

Security audit

agentmemory

Security checks for vulnerabilities and agentic risk

Overview

The plugin’s automatic memory capture and recall match its advertised purpose, with clearly disclosed privacy and sharing considerations.

Enable this only where you want conversations retained across sessions; do not assume sensitive text is automatically redacted. Set explicit preferences for unsolicited saves and recall. Use HTTPS for remote storage and enable optional MCP only for deliberately shared, trusted storage. Deleting saved notes does not erase captured conversations. The shipped native plugin was inspected directly, but the separate backend and other-host hook implementations were not available for verification.

Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Memory PoisoningPersistent Context Injection, Context Window Stuffing, Memory Manipulation
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (14)

Agent Config Directory Access

High
Category
Agent Snooping
Confidence
90% confidence
Finding

Skill reads from agent configuration directories (.claude/, .codex/, .gemini/). These directories may contain API keys, personal settings, and other credentials that the skill has no legitimate need to access.

Content

Scanner excerpt · skills/agentmemory-agents/REFERENCE.md (reported line 10)May include surrounding context.

md
| Agent | Name | Protocol |
| --- | --- | --- |
| Antigravity | `antigravity` | Using MCP via ~/.gemini/config/mcp_config.json, shared by current Antigravity surfaces. Pass --with-hooks for capture through ~/.gemini/config/hooks.json. |
| Antigravity CLI (agy) | `antigravity-cli` | Using MCP via ~/.gemini/config/mcp_config.json, shared with current Antigravity IDE versions. The `/mcp` slash command inside agy lists configured servers. Pass --with-hooks to also install the native ~/.gemini/config/hooks.json auto-capture hooks. |
| Claude Code | `claude-code` | Using MCP. Hooks are also available, see https://github.com/rohitg00/agentmemory. |
| Cline | `cline` | Using MCP via ~/.cline/data/settings/cline_mcp_settings.json (CLI), or CLINE_MCP_SETTINGS_PATH when set. VS Code users: add the same block via Cline Settings → MCP Servers → Edit JSON. |

Agent Config Directory Access

High
Category
Agent Snooping
Confidence
90% confidence
Finding

Skill reads from agent configuration directories (.claude/, .codex/, .gemini/). These directories may contain API keys, personal settings, and other credentials that the skill has no legitimate need to access.

Content

Scanner excerpt · skills/agentmemory-agents/REFERENCE.md (reported line 10)May include surrounding context.

md
| Agent | Name | Protocol |
| --- | --- | --- |
| Antigravity | `antigravity` | Using MCP via ~/.gemini/config/mcp_config.json, shared by current Antigravity surfaces. Pass --with-hooks for capture through ~/.gemini/config/hooks.json. |
| Antigravity CLI (agy) | `antigravity-cli` | Using MCP via ~/.gemini/config/mcp_config.json, shared with current Antigravity IDE versions. The `/mcp` slash command inside agy lists configured servers. Pass --with-hooks to also install the native ~/.gemini/config/hooks.json auto-capture hooks. |
| Claude Code | `claude-code` | Using MCP. Hooks are also available, see https://github.com/rohitg00/agentmemory. |
| Cline | `cline` | Using MCP via ~/.cline/data/settings/cline_mcp_settings.json (CLI), or CLINE_MCP_SETTINGS_PATH when set. VS Code users: add the same block via Cline Settings → MCP Servers → Edit JSON. |

Agent Config Directory Access

High
Category
Agent Snooping
Confidence
90% confidence
Finding

Skill reads from agent configuration directories (.claude/, .codex/, .gemini/). These directories may contain API keys, personal settings, and other credentials that the skill has no legitimate need to access.

Content

Scanner excerpt · skills/agentmemory-agents/REFERENCE.md (reported line 11)May include surrounding context.

md
| Agent | Name | Protocol |
| --- | --- | --- |
| Antigravity | `antigravity` | Using MCP via ~/.gemini/config/mcp_config.json, shared by current Antigravity surfaces. Pass --with-hooks for capture through ~/.gemini/config/hooks.json. |
| Antigravity CLI (agy) | `antigravity-cli` | Using MCP via ~/.gemini/config/mcp_config.json, shared with current Antigravity IDE versions. The `/mcp` slash command inside agy lists configured servers. Pass --with-hooks to also install the native ~/.gemini/config/hooks.json auto-capture hooks. |
| Claude Code | `claude-code` | Using MCP. Hooks are also available, see https://github.com/rohitg00/agentmemory. |
| Cline | `cline` | Using MCP via ~/.cline/data/settings/cline_mcp_settings.json (CLI), or CLINE_MCP_SETTINGS_PATH when set. VS Code users: add the same block via Cline Settings → MCP Servers → Edit JSON. |
| Codex CLI | `codex` | Using MCP. Hooks ship via the Codex plugin; on Codex Desktop, also pass --with-hooks to

Agent Config Directory Access

High
Category
Agent Snooping
Confidence
90% confidence
Finding

Skill reads from agent configuration directories (.claude/, .codex/, .gemini/). These directories may contain API keys, personal settings, and other credentials that the skill has no legitimate need to access.

Content

Scanner excerpt · skills/agentmemory-agents/REFERENCE.md (reported line 11)May include surrounding context.

md
| Agent | Name | Protocol |
| --- | --- | --- |
| Antigravity | `antigravity` | Using MCP via ~/.gemini/config/mcp_config.json, shared by current Antigravity surfaces. Pass --with-hooks for capture through ~/.gemini/config/hooks.json. |
| Antigravity CLI (agy) | `antigravity-cli` | Using MCP via ~/.gemini/config/mcp_config.json, shared with current Antigravity IDE versions. The `/mcp` slash command inside agy lists configured servers. Pass --with-hooks to also install the native ~/.gemini/config/hooks.json auto-capture hooks. |
| Claude Code | `claude-code` | Using MCP. Hooks are also available, see https://github.com/rohitg00/agentmemory. |
| Cline | `cline` | Using MCP via ~/.cline/data/settings/cline_mcp_settings.json (CLI), or CLINE_MCP_SETTINGS_PATH when set. VS Code users: add the same block via Cline Settings → MCP Servers → Edit JSON. |
| Codex CLI | `codex` | Using MCP. Hooks ship via the Codex plugin; on Codex Desktop, also pass --with-hooks to

Agent Config Directory Access

High
Category
Agent Snooping
Confidence
90% confidence
Finding

Skill reads from agent configuration directories (.claude/, .codex/, .gemini/). These directories may contain API keys, personal settings, and other credentials that the skill has no legitimate need to access.

Content

Scanner excerpt · skills/agentmemory-agents/REFERENCE.md (reported line 15)May include surrounding context.

md
re also available, see https://github.com/rohitg00/agentmemory. |
| Cline | `cline` | Using MCP via ~/.cline/data/settings/cline_mcp_settings.json (CLI), or CLINE_MCP_SETTINGS_PATH when set. VS Code users: add the same block via Cline Settings → MCP Servers → Edit JSON. |
| Codex CLI | `codex` | Using MCP. Hooks ship via the Codex plugin; on Codex Desktop, also pass --with-hooks to install the global hooks.json workaround. |
| Continue | `continue` | Using MCP via ~/.continue/config.yaml (preferred) or config.json (legacy, only when no yaml). |
| GitHub Copilot CLI | `copilot-cli` | Using MCP. Install the plugin too for full hooks/skills coverage. |
| Cursor | `cursor` | Using MCP. Lifecycle hooks ship in the Cursor plugin: Settings → Plugins → Add marketplace → this repo, then enable agentmemory. |
| Devin CLI | `devin` | Using MCP via the user config. Devin CLI migrates mcpServers into mcp_config.json on newer builds. Pass --with-hooks for native auto-c

External Transmission

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Content

Scanner excerpt · skills/agentmemory-rest-api/SKILL.md (reported line 20)May include surrounding context.

md
test -n "$agentmemory_secret" || exit 1

# save
curl -fsS -X POST http://localhost:3111/agentmemory/remember \
  -H "Authorization: Bearer $agentmemory_secret" \
  -H "Content-Type: application/json" \
  -d '{"content":"chose JWT refresh rotation","concepts":["jwt-refresh-rotation"]}'

Memory Manipulation

High
Category
Memory Poisoning
Confidence
80% confidence
Finding

Skill manipulates agent memory, state, or stored context. Memory corruption can alter personality, override safety rules, or cause unpredictable behavior.

Content

Scanner excerpt · skills/forget/SKILL.md (reported line 3)May include surrounding context.

md
---
name: forget
description: Delete selected durable memories from agentmemory after showing them and getting explicit confirmation. Use when the user says "forget this", "delete memory", "remove that note", or wants to scrub specific data for privacy.
argument-hint: "[what to forget - session ID, file path, or search term]"
user-invocable: true
---

Activation phrases overlap with ordinary questions that do not request memory retrieval.

Low
Category
Trigger Abuse
Confidence
96% confidence
Finding

The description explicitly activates the skill for phrases such as "did we ever" and "have we seen", which can occur in ordinary conversation without requesting stored session history. It does not qualify these phrases with a requirement for memory-search intent or provide exclusions, so unrelated questions could trigger retrieval.

Content

No source excerpt is available for this finding.

Everyday phrases such as "note that" can trigger persistent memory storage without a clear request to save information.

Medium
Category
Trigger Abuse
Confidence
94% confidence
Finding

The description treats ambiguous conversational phrases such as "note that" as requests for long-term storage, while the workflow directs the agent to save immediately without confirming persistence intent. This can cause incidental or sensitive information to persist beyond the current session and influence future retrieval without the user's informed consent. Memory storage is the skill's legitimate purpose, but the overly broad trigger creates an avoidable consent boundary failure.

Content

No source excerpt is available for this finding.

Standalone time expressions are listed as skill activation triggers.

Medium
Category
Trigger Abuse
Confidence
98% confidence
Finding

The description explicitly lists "today" and "this week" as reasons to activate the recap skill, although these expressions commonly occur in unrelated requests. It does not require an accompanying request for a project-session recap or provide exclusions, creating a risk of unintended invocation.

Content

No source excerpt is available for this finding.

The activation condition "starts a session with no fresh context" is ambiguous and can invoke session recovery without a request to resume prior work.

Medium
Category
Trigger Abuse
Confidence
96% confidence
Finding

The trigger "starts a session with no fresh context" permits retrieval and display of previous conversations without an explicit request to resume them. This can unexpectedly expose sensitive historical context in a fresh conversation or let stale context influence the new task. Directory-boundary checks and provenance validation reduce cross-project disclosure, but do not establish that recalling prior conversations is appropriate for the current interaction.

Content

No source excerpt is available for this finding.

The activation description includes the standalone word "always" and ordinary corrections without requiring an intent to save a lesson.

Medium
Category
Trigger Abuse
Confidence
94% confidence
Finding

The description treats common conversational wording such as "always" and "never do X" as activation conditions. Without requiring an explicit lesson-saving request or excluding task-local instructions, these triggers can invoke persistent lesson recording when the user only intended to direct the current task.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
94% confidence
Finding

The supplied code performs plugin installation rather than implementing, explaining, or diagnosing observation-capture hooks. Although the plugin name matches the described subject, installation is a substantive environment-changing action not disclosed by the description. The installed plugin's behavior cannot be determined from this command alone.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The supplied chunk performs plugin setup rather than explaining, debugging, or tuning automatic memory capture. Adding a marketplace and installing a plugin are materially different actions from the declared purpose. The chunk does not show lifecycle hooks or observation-capture behavior, and the installed plugin's behavior cannot be inferred from these commands alone.

Content

No source excerpt is available for this finding.

Endor dependency reachability

Checked Nov 8, 2026 · 9:48 PM UTC

No reachable vulnerabilities reported

Endor dependency scan completed.