Lesson 10: Performance, Cost, and Privacy Compliance

10.1 Token Cost Estimation

Utilizing memory systems incurs additional token overhead. Here are the estimated increments:

System	Startup Overhead	Runtime Overhead	Daily Token Increase (Est.)
auto memory	~5-25 KB	Nearly 0	~25 KB
claude-mem	~10-50 KB	"Distillation" cost post-tool call	~100-500 KB

Cost Estimate: Using claude-mem during active development typically adds approximately $0.05 - $0.30 to daily API costs (based on Claude 3.5 Sonnet).

10.2 Latency and Disk Usage

• Interaction Latency: claude-mem runs Hooks after every tool call, which can slow down operations by about 50-200ms. For batch script scenarios, using the --bare flag is recommended to improve speed.
• Disk Growth:
  • auto memory: Extremely slow growth, limited by the 25KB index constraint.
  • claude-mem: SQLite grows by ~10MB per 1,000 records; Chroma vector DB grows by ~20-50MB.

10.3 Privacy and Security Risks

Risk Point	auto memory	claude-mem
Code Upload Risk	Extremely Low (records key conclusions only)	Medium (distills nearly every tool interaction via remote LLM)
Secret Leakage Risk	Low (structured notes, human-auditable)	Medium (auto-captured, might distill sensitive info into Observations)
Cross-Project Leak	No (isolated by working directory)	Yes (shared globally; searches may retrieve info from other projects)

10.4 Security Recommendations for Sensitive Projects

For projects with strict privacy requirements:

1. Disable Plugins: Add claude-mem@thedotmack to disabledPlugins in the project's .claude/settings.local.json.
2. Offline Mode: Use claude --bare for tasks; this skips all Hooks, LSP, and memory systems.
3. Regular Audits:
   • Check ~/.claude/projects/*/memory/ for plaintext keys.
   • Audit the summary field in claude-mem.db via SQL: sqlite3 ~/.claude-mem/claude-mem.db "SELECT summary FROM observations" | grep -E 'sk-|pk_'