Alex Newman's Persistent Memory for Claude Code

Table of content

Alex Newman is a Philadelphia-based developer who spent 20 years building web products before turning his attention to AI tooling. His claude-mem plugin has 14.7k GitHub stars, making it one of the most popular Claude Code extensions. It solves a problem every Claude Code user knows: the AI forgets everything when you restart.

Background

20+ years as a web designer, developer, and product manager
Product Manager/Owner at Copter Labs (2015-2025), working with Intel, PlayStation, and Ashford University
Previously CTO at In The Rooms, a recovery social network
Runs @Claude_Memory on X, focused on AI memory systems
Open source contributor with 77 public repositories

GitHub | X | LinkedIn

The Memory Problem

Every Claude Code session follows the same pattern:

/init - Claude learns your codebase
Work for an hour - context fills up
/clear or restart - everything is gone
Next session - re-explain everything

The CLAUDE.md workaround helps but requires manual maintenance. Newman’s insight: use AI to compress AI’s own observations, then inject relevant context automatically.

How claude-mem Works

The plugin captures everything Claude does during a session, compresses it with AI using the Claude Agent SDK, and injects relevant context back into future sessions.

Install from Claude Code:

> /plugin marketplace add thedotmack/claude-mem
> /plugin install claude-mem

After restart, context from previous sessions appears automatically.

Architecture

Six core components handle the memory lifecycle:

Component	Purpose
5 Lifecycle Hooks	SessionStart, UserPromptSubmit, PostToolUse, Stop, SessionEnd
Worker Service	HTTP API on port 37777 with web UI
SQLite Database	Persistent storage for sessions and observations
Chroma Vector DB	Hybrid semantic and keyword search
mem-search Skill	Natural language queries with progressive disclosure
Smart Install	Dependency checker that caches prerequisites

The web viewer at localhost:37777 shows a real-time memory stream during sessions.

Progressive Disclosure

Newman’s approach to token efficiency uses a three-layer workflow for memory retrieval:

search      →  Compact index with IDs (~50-100 tokens/result)
timeline    →  Chronological context around results
get_details →  Full observation data (~500-1,000 tokens/result)

This saves approximately 10x tokens by filtering before fetching full details. Claude first searches, narrows down relevant observations, then retrieves only what it needs.

Privacy Controls

Sensitive content stays out of memory with <private> tags:

<private>
API_KEY=sk-secret-key
</private>

Everything between the tags gets excluded from compression and storage.

Configuration

Settings live in ~/.claude-mem/settings.json:

{
  "aiModel": "claude-sonnet-4-20250514",
  "workerPort": 37777,
  "dataDirectory": "~/.claude-mem/data",
  "logLevel": "info",
  "contextInjection": {
    "enabled": true,
    "maxTokens": 4000
  }
}

The Complexity Battle

Newman’s v8.1.0 release notes reveal a hard-won lesson about working with AI:

“For three months, Claude’s instinct to add code instead of delete it caused the same bugs to recur. What should have been 5 lines of code became ~1000 lines, 11 useless methods, and 7+ failed fixes. The real achievement: 984 lines of code deleted.”

The problem was Claude building redundant session management on top of session IDs that the hook system already provided. The fix: aggressive simplification.

Key Takeaways

Principle	Implementation
Compress, don’t transcribe	AI summaries beat raw logs for context retention
Progressive disclosure	Layer retrieval to balance completeness with token cost
Let hooks handle state	Don’t rebuild what the platform already provides
Delete more than you add	Simplicity beats features for memory systems
Privacy by exclusion	Explicit tags mark content to skip

Token Efficiency - Why compression matters for context windows
Three-Layer Workflow - The spec-implement-verify pattern claude-mem enables

Links

Next: John Lindquist’s TypeScript Hooks for Claude Code