Claude-Mem: Persistent Memory for AI Agents

Claude-Mem solves a fundamental problem: AI agents (including Claude Code) start fresh each session with no memory of what they've done before. Claude-Mem captures everything your agent does, compresses it intelligently, and injects relevant context back into future sessions—so your agent learns from its history.

AI-generated content: drafted with generative AI (Claude, by Anthropic), fact-checked and approved by a human editor. Full notice at the end of the article.

Source repository thedotmack/claude-mem ↗ 100,226 stars · JavaScript

The problem: stateless agents

When you run a Claude Code session or any AI agent, it has no built-in memory of previous interactions. Each new session is a blank slate. If your agent learned something useful last week—a workaround, a discovered API pattern, a user preference—that knowledge is lost.

For single-shot tasks this doesn't matter. But for ongoing automation, code review loops, or multi-stage projects, the agent repeats the same mistakes, rediscovers the same solutions, and wastes context tokens rehashing past events.

Claude-Mem fixes this by capturing session activity, compressing it with AI, and making it available as context for future runs. Your agent builds up a persistent knowledge base without you manually maintaining a changelog.

How Claude-Mem works

Claude-Mem is a Claude Code plugin built on lifecycle hooks. Per its README:

Capture: hooks on SessionStart, UserPromptSubmit, PostToolUse, Stop and SessionEnd record tool-use observations during each session.
Compress: a model turns those observations into semantic summaries.
Store: a local worker service (managed by Bun) keeps sessions, observations and summaries in SQLite with full-text search, plus a Chroma vector database for hybrid semantic + keyword search.
Retrieve: relevant context is injected into new sessions, and Claude can query past work through Claude-Mem's MCP search tools. A local web viewer shows your memory.

Besides Claude Code, the installer supports other agents and tools; check the README for the current list.

Fitting it into a Claude Code workflow

Claude Code runs discrete, often iterative tasks: debugging, generating features, refactoring. Without memory, you repeat context: "I tried approach X last session and hit problem Y." Claude-Mem eliminates that friction.

A typical workflow:

You start a Claude Code session to work on a feature.
Claude-Mem captures the agent's reasoning, code decisions, test results, and errors.
Next session, Claude-Mem injects a summary: "Previously discovered that auth library version 2.x breaks with async patterns; use 1.9.x instead."
Claude Code uses this to avoid the same pitfall.

This is especially valuable for long-running automation tasks where the agent needs to adapt based on past attempts. It also reduces context bloat—instead of pasting full logs, you get a compressed memory.

Real limitations

Hosted by default: the installer asks you to sign in, and the default memory is the hosted CMEM Pro service. If your code or prompts must stay on your machine, opt out: pass an explicit --provider flag or set CLAUDE_MEM_ONLINE_OPTIN=false, as the README describes.
Compression costs tokens: summarising observations calls a model, which adds cost and some latency.
Summaries lose detail: for exact past outputs (an exact error message, a full diff) keep using git history and logs.
Hooks run on every session: review what a plugin's hooks do before installing, and expect memory to grow over time.
Stale memories: a lesson from last month can be wrong after a refactor; prune memory that no longer applies.

Getting started

Install (sets up the plugin, hooks and worker)
npx claude-mem install

Or from inside Claude Code:

Install from the plugin marketplace
/plugin marketplace add thedotmack/claude-mem
/plugin install claude-mem

Don't use npm install -g claude-mem for this: the README notes it installs only the SDK/library and does not register the hooks or start the worker. Restart Claude Code after installing; the next sessions start building memory automatically. Claude-Mem is Apache-2.0 licensed.