Audit candidate: MemPalace/mempalace

[xiaolai]

Repo

https://github.com/MemPalace/mempalace

Field	Value
Stars	51982
Source	manual queue (2026-05-12 discovery batch)
Description	The best-benchmarked open-source AI memory system. And it's free.

Next Steps

1. Review the repo — check CONTRIBUTING.md, recent activity, PR history
2. If worthy: add label audit-ready to trigger the audit workflow
3. If not: close this issue

Checklist

• Accepts contributions
• Updated in last 60 days
• Not a fork or mirror
• Has meaningful NL artifacts

[github-actions]

Audit Complete: MemPalace/mempalace

NL Score: 90/100 | Security: CLEAR | Artifacts: 15 | Strategy: single

NLPM Audit: MemPalace/mempalace

Date: 2026-05-13 | Artifacts: 15 | Strategy: single
NL Score: 90/100
Security: CLEAR
Bugs: 1 | Quality Issues: 13 | Security Findings: 1

NL Score Summary

File	Type	Score	Top Issue
.claude-plugin/commands/mine.md	command	70	Multi-step body, no output format, no empty input handling
.claude-plugin/commands/init.md	command	80	Multi-step body lacks numbered steps, no output format
.claude-plugin/commands/search.md	command	80	No output format, no empty input handling
.claude-plugin/plugin.json	manifest	85	commands: [] while 5 command files exist
.claude-plugin/commands/help.md	command	90	No output format specified
.claude-plugin/commands/status.md	command	90	No output format specified
.claude-plugin/skills/mempalace/SKILL.md	skill	94	Vague: "short", "often", "significant"
integrations/openclaw/SKILL.md	skill	94	Stale version (3.3.0 vs 3.3.5); vague quantifiers
CLAUDE.md	project-doc	94	Vague: "real understanding", "instant", "vast amounts"
.claude-plugin/hooks/hooks.json	hook-config	95	Clean
.codex-plugin/skills/help/SKILL.md	skill	95	Clean
.codex-plugin/skills/init/SKILL.md	skill	95	Clean
.codex-plugin/skills/mine/SKILL.md	skill	95	Clean
.codex-plugin/skills/search/SKILL.md	skill	95	Clean
.codex-plugin/skills/status/SKILL.md	skill	95	Clean

Security Scan

Severity	Count
Critical	0
High	0
Medium	0
Low	1

Execution Surface Inventory

Surface	Files
Hook scripts (root)	hooks/mempal_save_hook.sh, hooks/mempal_precompact_hook.sh
Hook scripts (claude-plugin)	.claude-plugin/hooks/mempal-stop-hook.sh, .claude-plugin/hooks/mempal-precompact-hook.sh
Hook scripts (codex-plugin)	.codex-plugin/hooks/mempal-hook.sh
MCP config	.claude-plugin/.mcp.json
Package manifest	website/package.json

Security Findings

#	Severity	File	Line	Pattern	Description
1	Low	website/package.json	null	SEC-unpinned-semver	Five dev dependencies use caret ranges (^1.8.0, ^11.14.0, ^1.6.4, ^2.0.17, ^3.5.32); private docs package so supply-chain exposure is minimal

Hook review notes: The two root hooks (mempal_save_hook.sh, mempal_precompact_hook.sh) demonstrate deliberate security hardening: they use mapfile + single-pass Python sanitization to avoid eval, validate transcript paths for .json/.jsonl suffix and absence of .. traversal, and pass the transcript file via sys.argv to the Python inline script rather than interpolating into the command string. The .claude-plugin and .codex-plugin wrapper hooks are thin CLI delegates with no logic of their own. No critical or high patterns detected.

Bugs (PR-worthy)

---

Full report: https://github.com/xiaolai/nlpm-for-claude/blob/main/auditor/audits/MemPalace-mempalace.md

2 bug(s) found. To submit fix PRs to the target repo, add the contribute-approved label to this issue.

[github-actions]

No high-confidence findings — no PRs opened

The audit produced no findings at confidence: high.
Medium / low / unspecified-confidence findings are kept in
the audit data for our own learning but never become PRs.
This keeps maintainers free of unverified suggestions.

[github-actions]

No high-confidence findings — no PRs opened

The audit produced no findings at confidence: high.
Medium / low / unspecified-confidence findings are kept in
the audit data for our own learning but never become PRs.
This keeps maintainers free of unverified suggestions.

[github-actions]

No high-confidence findings — no PRs opened

The audit produced no findings at confidence: high.
Medium / low / unspecified-confidence findings are kept in
the audit data for our own learning but never become PRs.
This keeps maintainers free of unverified suggestions.