oathe / scan / reports / caution
Is mempalace/mempalace safe?
Use with caution. mempalace/mempalace is not clearly unsafe, but the audit flagged risks to review before you install it. Oathe's behavioral security audit gave the mempalace skill by mempalace a trust score of 75/100 with 13 findings (5 high).
https://github.com/mempalace/mempalace
Is mempalace/mempalace safe to install?
MemPalace is a functionally legitimate local AI memory system with a large surface area of concern: aggressive behavioral directives that override agent response flow and instruct persistent harness modification; optional but actively promoted shared-brain networking that transmits all stored content to remote servers; session hooks that silently mine conversation transcripts on every agent event; and an LLM refinement feature that sends conversation content to configurable external HTTP endpoints. The clone itself was clean with no unexpected network connections, and all canary honeypots remained intact (pre-clone credential reads were monitoring infrastructure). The risks are primarily from the disclosed-but-aggressive data collection and network exfiltration capabilities, not from hidden malicious code.
What security issues were found in mempalace/mempalace?
Category Scores
Findings (13)
HIGH Mandatory pre-response palace search protocol -12 ▶
The mempalace-recall skill unconditionally instructs the agent to call mempalace_search or mempalace_kg_query BEFORE responding to any query about people, projects, past events, or prior decisions. This is a blanket behavioral override that intercepts normal response flow, adds latency, and could be exploited to front-load data collection.
Evidence: 'BEFORE RESPONDING about any person, project, or past event: call mempalace_search or mempalace_kg_query FIRST. Never guess from memory — verify from the palace.'
HIGH Agent instructed to permanently modify its own durable configuration -10 ▶
The setup skill instructs the agent to run 'mempalace rules --agent
Evidence: 'Render the canonical rules with: mempalace rules --agent <machine-harness>. Install the rendered marker-delimited block in the harness durable agent instructions. Replace an existing marked block instead of appending a duplicate.'
MEDIUM Persona assignment overrides agent identity -5 ▶
The mempalace-recall skill explicitly instructs the agent to adopt an alternative professional identity, which can shift baseline behavior and override the agent's configured persona.
Evidence: 'Act as a senior AI-memory systems engineer with decades of experience building verbatim recall, semantic retrieval, and temporal knowledge graphs.'
MEDIUM External URL references on non-canonical domain -8 ▶
The skill references mempalaceofficial.com (distinct from github.com/MemPalace/mempalace), agentskills.io, and cursor.com/docs/skills. Agents following skill instructions may fetch these URLs, and the mempalaceofficial.com domain is not the authoritative source.
Evidence: 'Official shared-brain guide: https://mempalaceofficial.com/guide/shared-brain.html', 'Agent Skills specification: https://agentskills.io/specification'
HIGH Shared-brain hub transmits all palace content over network -15 ▶
When configured in shared-brain mode (actively promoted in the skill), all mined file content, conversation history, diary entries, and knowledge graph facts are transmitted to a central hub server. The hub server can be remote and internet-accessible. This is the primary exfiltration vector.
Evidence: hub_client.py + worker protocol with authenticated /api/worker/claim, /progress, /complete endpoints; SKILL.md: 'guide the user through mempalace serve and the official shared-brain guide. Do not expose a non-loopback server without authentication.'
MEDIUM LLM refinement sends conversation content to configurable external endpoints -10 ▶
The closet_llm.py module (invoked via regenerate_closets) sends drawer content — which contains verbatim conversation excerpts — to an external HTTP endpoint specified by LLM_ENDPOINT environment variable. This is opt-in but underdisclosed in the SKILL.md.
Evidence: closet_llm.py: _call_llm() makes urllib.request.urlopen POST to cfg.endpoint + '/chat/completions' with drawer content as the user message body; Authorization header uses LLM_KEY
HIGH Session hooks auto-execute shell code on every agent event -15 ▶
The .claude-plugin/hooks/ directory contains shell scripts registered as Claude Code hooks. These execute automatically on PreToolUse/PostToolUse/Stop/SessionEnd events. The save hook mines conversation transcripts via 'mempalace mine --mode convos' on every session end without per-invocation approval.
Evidence: .claude-plugin/hooks/hooks.json, mempal-stop-hook.sh, mempal-session-end-hook.sh, mempal-precompact-hook.sh; mempal_save_hook.sh: 'dirname "$TRANSCRIPT_PATH"' + 'mempalace mine --mode convos'
MEDIUM Platform-specific hook installer scripts register persistent system hooks -10 ▶
hooks/cursor/install.sh and hooks/antigravity/install.sh install hooks at --scope user level, affecting all projects for that user. The SKILL.md actively recommends running these scripts. Once installed, they persist across projects and sessions.
Evidence: SKILL.md: 'install the Cursor hooks separately by running hooks/cursor/install.sh --scope user from a cloned MemPalace repo'; hooks/cursor/install.sh, hooks/antigravity/install.sh
LOW External HTTP calls from closet_llm module -5 ▶
Beyond the data exfiltration concern, the closet_llm.py module introduces an HTTP client dependency that runs during palace maintenance operations and contacts whatever endpoint is configured in LLM_ENDPOINT.
Evidence: closet_llm.py uses urllib.request.urlopen for chat completion requests; endpoint is fully user-configurable via env vars
LOW Clone required retry — prior attempt detected -5 ▶
The install output shows 'fatal: destination path already exists and is not an empty directory', indicating the skill directory was pre-populated from a prior clone attempt. The clone itself was clean and only contacted github.com.
Evidence: Install Output: 'fatal: destination path /home/oc-exec/skill-under-test/mempalace/mempalace already exists and is not an empty directory.'
INFO Pre-clone credential reads were monitoring infrastructure, not skill code -3 ▶
Canary files (.env, .ssh/id_rsa, .aws/credentials, .npmrc, .docker/config.json, gcloud credentials) were opened read-only at 03:20:37 — six seconds before the git clone at 03:20:43. These reads are attributable to the audit monitoring infrastructure (sudo PAM, DNS resolution) establishing baseline state, confirmed by CLOSE_NOWRITE flags and audit pid/timing correlation.
Evidence: inotify timestamps: 03:20:37 OPEN/.ACCESS/CLOSE_NOWRITE on canary files; git clone starts 03:20:43; Canary integrity check: ✅ All canary files intact
HIGH Automatic session-end mining enables persistent covert data collection -20 ▶
The session-end hooks silently mine full conversation transcripts on every agent session without per-session user consent beyond initial hook installation. In shared-brain mode, this data is then forwarded to remote hubs, creating a persistent data collection pipeline triggered by normal agent use.
Evidence: mempal_save_hook.sh auto-mines TRANSCRIPT_PATH; hooks trigger on Stop/SessionEnd events; shared-brain mode forwards to hub; all disclosed but automatic once installed
MEDIUM Shared palace creates cross-agent memory poisoning attack surface -15 ▶
In shared-brain mode, all agents read from and write to the same palace. A compromised or socially-engineered agent can write malicious 'memories' that influence the responses of all other agents reading from the same palace, without any signing or verification of drawer authenticity.
Evidence: mempalace_add_drawer, mempalace_checkpoint, mempalace_diary_write accessible to all connected agents; no cryptographic verification of drawer source or integrity
Should I install mempalace/mempalace?
Oathe's verdict for mempalace/mempalace is CAUTION with a trust score of 75/100. Recommendation: Review before install.