Claude-Mem Saves 95% Tokens and Offers Unlimited Memory – 25.8K‑Star GitHub Project

Pain Point: AI Coding Assistants Forget

Developers experience repeated context loss when using AI coding assistants such as Claude Code; after a session ends the assistant no longer remembers project structure, recent bug fixes, or design decisions, forcing a costly re‑explanation of the codebase.

Innovation: Three‑Layer Progressive Disclosure

Claude‑Mem introduces a three‑stage retrieval workflow that only loads detailed information when needed, unlike traditional approaches that dump the entire history into the context window.

Layer 1 – Index View: On session start, Claude‑Mem presents a concise index containing recent session titles, types (bug‑fix or feature), timestamps, and estimated token cost. This step consumes roughly 50–200 tokens.

Layer 2 – On‑Demand Query: A natural‑language query triggers the MCP search tool to fetch relevant index entries, consuming about 100–500 tokens per entry.

Layer 3 – Full Detail: When deeper insight is required, the full observation is retrieved, costing 500–1000 tokens per entry.

This architecture reduces a typical 20‑record request from 10,000–20,000 tokens to 2,500–5,000 tokens, saving about 95% of token usage.

Core: AI Semantic Compression

Claude‑Mem uses the Claude Agent SDK to compress raw tool outputs (often >50,000 tokens per session) into structured memories of 100–500 tokens, achieving compression ratios of 10:1 to 100:1.

Each compressed observation stores six fields: title, subtitle, narrative, facts, concepts, and type. Example of raw output:

[Read worker-service.ts]
[Read hooks.ts]
[Bash test]
[Edit worker-service.ts] 添加 await before process.spawn()

Compressed representation:

Title: Fixed worker startup race condition
Type: bugfix
Concepts: [concurrency, startup, timing]
Facts:
- Worker process spawned before port check completed
- Added await before process.spawn()

Visualization: Transparent Memory Flow

A web viewer UI (React + TypeScript, SSE‑driven) displays the real‑time memory stream, allowing developers to verify captured observations, filter by project, and persist settings in the browser.

Search: Hybrid Retrieval Architecture

Claude‑Mem combines SQLite FTS5 keyword search (≈12 ms latency) with optional ChromaDB semantic search. Queries first narrow results via FTS5, then enrich with vector embeddings. If ChromaDB fails, the system gracefully falls back to pure FTS5, ensuring robustness.

Effect: Real‑World Performance Gains

Compared with earlier v3, Claude‑Mem v5 reduces per‑session token consumption from 25,000 to 1,100 (96% reduction), raises relevant context coverage from 8% to 100%, and cuts hook execution time from 200 ms to 10 ms. Tool‑call limits increase twenty‑fold, enabling far more operations per session.

User feedback highlights dramatic time savings (≈30 minutes per session) and continuous decision‑making context, turning the assistant into a long‑term partner rather than a transient helper.

Impact: Workflow Transformation

By eliminating context‑switch costs, Claude‑Mem changes how developers collaborate with AI, automating knowledge capture, enabling “archaeology” of code evolution, and supporting team‑wide memory sharing.

Future: Towards Retrieval‑Augmented Decisions (RAD)

Claude‑Mem exemplifies a shift from Retrieval‑Augmented Generation (RAG) to Retrieval‑Augmented Decisions (RAD), emphasizing agent‑generated memory, adaptive indexing, and semantic ranking. The project roadmap includes adaptive index sizing, cross‑project memory, and collaborative sharing.

Getting Started: Two‑Command Installation

In a Claude Code session, run:

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

After restarting Claude Code, the memory system activates automatically. Configuration resides at ~/.claude-mem/settings.json. Sensitive content can be excluded using the <private> tag.

Summary

Claude‑Mem’s three‑layer progressive disclosure and AI semantic compression solve the cross‑session memory loss of AI coding assistants, cutting token usage by 95% and boosting tool‑call capacity twenty‑fold, as evidenced by extensive benchmark data and strong community adoption (25.8K+ GitHub stars).