If you use Claude Code daily, you know this frustration. You spent two hours explaining your architecture, fixing a complex bug, defining your code conventions. You close the session. The next day, you relaunch Claude Code, and it remembers nothing. Not the bug you fixed. Not the library you explicitly rejected. Not your project structure.

Developers call this "session amnesia." Every new conversation starts with a blank context window. Claude Code becomes a brilliant stranger who needs to rediscover everything.

The concrete result: you lose between 10 and 30 minutes at the start of each session rebuilding context. Over a work week, that adds up to several hours of lost productivity.

This is precisely the problem claude-mem aims to solve.

claude-mem is a free, open-source plugin for Claude Code, created by Alex Newman (@thedotmack). Its purpose is simple: give Claude Code persistent memory across sessions.

In practice, the plugin automatically captures everything Claude does during a coding session (file modifications, executed commands, architectural decisions, fixed bugs), compresses that information, and reinjects it into subsequent sessions. No more re-explaining anything.

The promise: transform Claude Code from a "competent developer who starts from scratch" into a "senior developer who remembers your codebase."

The project has 37.2k GitHub stars, 174 releases, and 22 contributors since its publication in August 2025. These numbers make it one of the most popular plugins in the Claude Code ecosystem.

The system operates on a fully automatic five-step cycle, completely transparent to the user:

1. Session start: claude-mem queries its database and injects relevant summaries from previous sessions into the context window.

2. During the session: every Claude action (file edits, bash commands, searches) is captured in real time and stored.

3. Session end: the plugin generates compressed semantic summaries of everything that happened.

4. Next session: only relevant context is reinjected, not the entire history.

You don't have to do anything. No copy-pasting, no special prompts, no files to maintain manually. The plugin handles everything in the background.

Behind this apparent simplicity, claude-mem uses a two-layer storage architecture:

• SQLite with FTS5: fast keyword-based search. Stores sessions, individual observations, and AI-generated summaries.

• Chroma (vector database): semantic search. Finds conceptually related context even without exact keyword matches. For example, a search for "authentication issues" will find JWT-related context even if the word "authentication" doesn't appear in the stored observations.

This hybrid approach combines the best of both worlds: the speed of exact-match search and the intelligence of semantic search.

To save tokens (and therefore costs), claude-mem uses a progressive disclosure system:

1. Search: returns a compact index with identifiers (about 50 to 100 tokens per result)

2. Timeline: shows context around a specific observation

3. Full details: retrieves complete information only for relevant items (about 500 to 1,000 tokens per result)

This system saves roughly 10x more tokens compared to loading all observations at once. Fewer tokens consumed means longer, more efficient sessions.

Installation takes under 30 seconds:

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

Restart Claude Code. Done. No API key required, no configuration files to create. Context from previous sessions automatically appears in new sessions.

Prerequisites: Node.js 18.0.0 or higher and the latest version of Claude Code with plugin support. Bun (JavaScript runtime) and uv (Python package manager) are installed automatically if missing.

Community feedback is largely positive, though with some nuance. On Reddit, a dedicated post received 729 upvotes, with comments describing the concept as "ingenious" and praising the plugin's flexibility on large codebases.

Developers using similar persistent memory tools report concrete improvements:

• Roughly 60% reduction in time spent re-explaining concepts between sessions

• Generated code follows project conventions about 85% of the time (versus 30% without persistent memory)

• Debugging is more precise because Claude recognizes recurring patterns

• Time from session start to working feature is cut roughly in half

These figures come from a developer using Mem0 (a competing tool), but the gains are representative of what persistent memory brings to any Claude Code workflow.

claude-mem proves most valuable in several scenarios:

• Long-running projects: on multi-week developments, accumulated context becomes a significant advantage. Claude remembers architecture choices, tested and rejected libraries, adopted patterns.

• Regression debugging: Claude can reference the exact change that introduced a bug, even if it happened several sessions ago.

• Multi-session refactors: maintain coherent architectural understanding across days of work.

• Preference learning: Claude automatically learns your style (arrow functions, TypeScript patterns, test conventions).

The market for AI coding assistant memory solutions has grown rapidly. Here is how claude-mem stacks up against the main alternatives:

claude-mem vs Mem0: claude-mem is entirely local (your data never leaves your machine), while Mem0 primarily operates in the cloud. claude-mem uses hooks (automatic, no agent action required), whereas Mem0 uses the MCP protocol (the agent must explicitly call memory tools).

claude-mem vs Supermemory: both use the Claude Code plugin system, but Supermemory requires a paid subscription. claude-mem is and remains free.

claude-mem vs native memory: Claude Code natively offers CLAUDE.md files (written by the developer) and MEMORY.md (written by Claude, limited to 200 lines). claude-mem complements this system by adding a large-scale recall layer with hybrid search. The two approaches are complementary, not competing.

claude-mem is distributed under the AGPL-3.0 license. Using it as-is is fine. However, if you modify the code and deploy that modified version as a network service, you must publish your source code. This is worth considering for companies looking to integrate it into a commercial product.

Also note: the ragtime/ directory is under a separate license (PolyForm Noncommercial License 1.0.0), restricted to personal and non-commercial use only.

claude-mem is not a perfect solution. Here is what to know before adopting it:

Technical limitations:

• chroma-mcp process leaks have been reported: orphaned subprocesses accumulate over time, causing SQLite lock errors on the vector database. Manual restart is sometimes needed.

• The tool is not compatible with Windows (Claude Code hooks freeze the CLI).

• Occasional crashes have been reported on macOS.

Conceptual limitations:

• Like any RAG (Retrieval-Augmented Generation) system, search quality can degrade as the database grows.

• Compressed logs don't interconnect: unlike a knowledge graph (like A-MEM), memories remain isolated from each other. Newer information doesn't update older memories.

• The plugin is designed for individual use. There is no shared memory across multiple developers working on the same project.

The Endless Mode case: this beta feature promises up to 95% token reduction and roughly 20x more tool calls before hitting context limits. However, these figures were disputed by the community during December 2025 discussions, with some calling them "exaggerated" for a then non-functional feature. The standard mode delivers more modest but real savings.

If you use Claude Code daily for projects spanning several days or weeks, claude-mem brings a tangible improvement to your workflow. Installation is trivial, the cost is zero, and the time saved at the start of each session is real.

For developers who primarily work on short, one-off tasks, the benefit will be smaller. Claude Code's native memory (CLAUDE.md + MEMORY.md) may suffice.

For companies, check the AGPL-3.0 license implications before integrating the plugin into your internal workflows. And keep an eye on cloud alternatives like Mem0 if large-scale semantic search is a priority.

The project is actively maintained (174 releases in 7 months), the community is large (37.2k GitHub stars), and the tool addresses a real need. Session amnesia remains one of the major friction points of current AI coding assistants, and claude-mem is today the most mature open-source solution for addressing it.