Claude-Mem
Persistent memory compression for Claude Code and other agent IDEs. Captures session observations, compresses them with AI, and injects relevant context into future sessions. ~85K GitHub stars.
What It Is
Claude-Mem is a hook-based memory system built for Claude Code Harness workflows. A dedicated observer captures tool usage and decisions during sessions, generates semantic summaries, and surfaces them in later sessions so agents do not re-derive project context from scratch. Source: github.com/thedotmack/claude-mem, 2026-06-01
Install: npx claude-mem install or via Claude Code plugin marketplace (/plugin marketplace add thedotmack/claude-mem). Also supports Gemini CLI, OpenCode, OpenClaw, Codex, and Copilot per upstream README. Source: github.com/thedotmack/claude-mem, 2026-06-01
Claimed Benefits
Community signal (not independently benchmarked in this wiki): up to 95% fewer tokens per session and ~20× more tool calls before context limits, by replacing full session replay with compressed retrieval. Source: X/@RoundtableSpace, 2026-03-15
Upstream positions Claude-Mem as complementary to RAG (external knowledge) vs RAD (agent working memory); a proposed open standard for hook-based agent memory. Source: claude-mem.ai, 2026-06-01
Source Snapshot
As of the 2026-07-01 registry refresh, npm claude-mem is 13.9.2 (latest), Git tag v13.9.2 resolves to 7ae038e22123d5545128459ab5bf23d7df16d7fc, and default-branch HEAD resolves to cbdce2d676a9646b73dc236eebb53e4019870dcf. The package requires Node >=20.12.0 or Bun >=1.0.0. The 2026-06-30 snapshot recorded 85,143 GitHub stars and Apache-2.0 licensing. Brin scans returned caution for both repo and skill surfaces, with deep scans pending. Keep it as a reference/candidate until a dedicated security audit approves the hook/plugin surface. Source: npm registry packument; git ls-remote, 2026-07-01; GitHub API and Brin API, 2026-06-30
Relationship to Kevin's Wiki
Kevin's wiki already implements compiled truth + timeline separation as durable memory. Claude-Mem is the session-level analogue for coding agents: procedural capture during work, compression, retrieval on next session. Pairs with Lean Ctx (token reduction via trimming) and The Brain-Agent Loop (wiki as long-horizon brain). Source: compiled from brain-agent-loop.md, lean-ctx.md
The caveman output-compression pattern referenced in Claude Code Harness is now tracked through Caveman Token Skill (JuliusBrussee/caveman), not as part of the Claude-Mem package itself. Source: Claude Code Harness; GitHub API, 2026-07-03
Timeline
- 2026-07-01 | Refreshed package/source snapshot:
claude-mem@13.9.2, tagv13.9.2at7ae038e22123d5545128459ab5bf23d7df16d7fc, default HEADcbdce2d676a9646b73dc236eebb53e4019870dcf. Source: npm registry packument;git ls-remote, 2026-07-01 - 2026-06-30 | Deep-reviewed the original Roundtable bookmark, local screenshot, npm package, GitHub source, and Brin scan. Current version is
13.9.1; Brin remains caution/pending, so no executable skill import. Source: X/@RoundtableSpace, 2026-03-15; Source: npm/GitHub/Brin, 2026-06-30 - 2026-06-24 | Confirmed this remains a tool/reference page, not an executable skill catalog page; removed the duplicate
wiki/skills/claude-mem.mdorphan. Source: User request, 2026-06-24 - 2026-06-01 | Enriched absorption: resolved full README + claude-mem.ai marketing copy. Dedicated tier-1 tool page created. Source: compiled from enrichment pipeline, 2026-06-01
- 2026-03-15 | @RoundtableSpace viral bookmark: "INFINITE MEMORY FOR FREE"; 1,407 likes, 2,588 bookmarks. Claims 95% token reduction, 20× tool calls. Source: X/@RoundtableSpace, 2026-03-15