Sign In
← Back to blog

Best Memory Skills for Claude Code (Repo-Checked)

2026-10-26•7 min
MemoryClaude CodeCursor

The short answer: there is no single best memory skill for Claude Code. The right pick depends on where you want memory to live. For notes that survive between sessions, a file-based skill such as Personal Assistant or Self-Reflect is enough. For semantic recall across tools, look at OpenViking or memU.

Every fact here comes from a static read of each repo at the commit named, on 2026-10-26. Nothing was run. Install steps are what each README says, not confirmed on your machine.

What "memory" means for a coding agent

Three different things are sold under the same word:

| Type | Where memory lives | Typical cost | Main risk |

|---|---|---|---|

| Notes files | Markdown such as CLAUDE.md, TASKS.md, a memory/ folder | None, no server | Grows stale; you maintain it |

| Hooks plus a local store | Hooks capture sessions into files or SQLite on your machine | Low | Raw transcripts land on disk |

| Memory server or hosted service | A database your agent queries over the network | Embedding or model keys, a server to run | Your code context leaves the machine |

Claude Code already loads CLAUDE.md at session start, so the cheapest memory is a tidy file. The tools below add capture, search or sync on top.

The tools, from the repos

Everything Claude Code (now published as ECC)

  • Repo: affaan-m/ECC (the older name was everything-claude-code). Commit 25278e7, 2026-10-26. License: MIT.
  • What it is: a large harness, not a memory tool: agents, skills, hooks, rules, "memory, continuous learning" and a security scanner in one package.
  • How it stores memory: lifecycle hooks (SessionStart, PreCompact, Stop) save and reload session summaries as files under your Claude config folder. The hook README says to "keep persistence local by default", and ECC_SESSION_START_CONTEXT=off disables the session-start context. An opt-in MCP memory vault (ecc-memory-mcp) is described separately.
  • Install for Claude Code (README): /plugin marketplace add https://github.com/affaan-m/ECC then /plugin install ecc@ecc, or npx ecc-universal@2.2.3 install --guided. It also lists Antigravity, Cursor and Gemini CLI targets, labelled capability-limited adapters.
  • Keys and network: none stated for the hooks; Node.js 18+ is required.
  • Best for: people who want a full setup and accept it installs far more than memory.

    • OpenViking

  • Repo: volcengine/OpenViking. Commit 5c066d9, 2026-10-26. License: AGPL-3.0.
  • What it is: a context database: a virtual filesystem under viking:// for memory, resources and skills, with generated directory summaries.
  • How it stores memory: on an OpenViking server. The Claude Code plugin recalls before each prompt and captures after each turn via hooks.
  • Install for Claude Code (README): run a server first (uv tool install openviking --upgrade && openviking-server init), then curl -fsSL https://openviking.ai/install | bash and pick Claude Code.
  • Keys and network: the README says a self-hosted server needs "a model provider with an embedding model and a VLM", so you supply an API key or a local model. The installer needs Node.js 18+ and macOS or Linux.
  • Best for: teams wanting one inspectable store shared by several agents. AGPL matters if you modify and host it.

    • MemOS

  • Repo: MemTensor/MemOS. Commit adcb9f2, 2026-10-23. License: Apache-2.0.
  • What it is: a memory framework and API for LLM apps, with four entry points: hosted Cloud API, Docker self-host (Neo4j and Qdrant per its table), and cloud or local plugins for OpenClaw, Hermes and DeepSeek Harness.
  • Claude Code route: the README documents none, so we do not treat it as a drop-in memory skill for Claude Code or Cursor.
  • Keys and network: the Cloud API needs a vendor key; self-hosting needs your own model keys in .env.
  • Best for: developers adding memory to their own application.

    • Cipher, now ByteRover CLI

  • Repo: campfirein/cipher, whose README is titled "ByteRover CLI" (brv). Commit 1052ac1, 2026-06-26, the most recent commit in the default branch at the time of reading. License: Elastic License 2.0.
  • What it is: a CLI with a curated context tree, web dashboard, git-like versioning and an MCP server (brv mcp). The README says it works with "22+ AI coding agents (Cursor, Claude Code, Windsurf, Cline, and more)".
  • Install (README): curl -fsSL https://byterover.dev/install.sh | sh or npm install -g byterover-cli, then run brv in your project. Per-agent steps live in external docs we did not read.
  • Keys and network: it runs on an LLM provider of your choice, and the README says "everything works locally by default". An optional cloud sync and hosted LLM require an account.
  • Best for: individuals wanting curated, reviewable project knowledge. Check the licence before building a service on it.

    • memU

  • Repo: NevaMind-AI/memU. Commit 42ef9fa, 2026-10-16. License: Apache-2.0.
  • What it is: a Python package (memu-cli) that mines your agent's session logs into a shared memory store and adds a retrieve-first instruction to your instruction file.
  • How it stores memory: the README lists the Claude Code adapter (memu-claude-code) as reading ~/.claude/projects//.jsonl and editing ~/.claude/CLAUDE.md. Self-hosted memory lives in SQLite (typically ~/.memu/memu.sqlite3) or Postgres; the cloud option stores it at memu.so.
  • Install for Claude Code (README): ask your agent to read the repo's SKILL.md and follow it. That file runs pip install --upgrade memu-cli, then prints the guide for your host.
  • Keys and network: local mode says "Embedding key required"; cloud mode needs a memu.so API key.
  • Best for: people using several agents who want one shared memory. It reads full transcripts; see privacy below.

    • Memory skills we already reviewed

  • Personal Assistant is Anthropic's productivity plugin from anthropics/knowledge-work-plugins, reviewed at commit 8444efc. It keeps a TASKS.md list and a two-tier memory: CLAUDE.md as working memory plus a memory/ folder. Our review summary: "Low risk: Markdown-only skills that write local task and memory files; external access only through MCP connectors you authorise."
  • Self-Reflect is the claude-reflect plugin, reviewed at commit b6c4232. Hooks queue corrections like "no, use X", and /reflect writes approved ones into CLAUDE.md or AGENTS.md. Our review summary: "Local Python hooks plus a claude -p call; it writes rules into your CLAUDE.md/AGENTS.md only after you approve them." In Antigravity only its short SKILL.md loads; the learning loop does not run.

    • How to choose

  • You want the least moving parts: start with notes files and a reviewed skill like Personal Assistant. No server, nothing leaves your machine.
  • You keep repeating the same corrections: Self-Reflect, because it asks before writing.
  • You want search across projects and tools: OpenViking or memU, and accept a store plus an embedding key.
  • You want one big harness: Everything Claude Code, with hooks you have read.
  • You build agent apps: MemOS.

    • On Antigravity, hooks and slash commands usually do not carry over, so file-based skills travel best. See how to install skills in Antigravity, our reviews and the awesome Antigravity skills list. New to the format? Read what are AI agent skills.

    Privacy notes

  • Memory files can hold secrets. Session logs capture whatever you pasted or the agent printed, including tokens and .env contents. Keep memory folders out of git unless reviewed.
  • Know where capture happens. memU's adapter reads full Claude Code session files. OpenViking's plugin sends turns to its server. Both are reasonable designs, but the destination decides the risk.
  • Prefer local first. Where a tool offers local and cloud modes, start local.
  • Check licences. AGPL, Elastic and MIT differ in what you may do with the code.
  • Hooks execute code. Read hook scripts before enabling them.

    • FAQ

    What is the best memory skill for Claude Code?

    It depends on whether you want files, a local store or a server. A small file-based setup is the safest start.

    Do these work in Antigravity or Cursor?

    Partly. ECC and OpenViking list Cursor adapters, and ECC lists an Antigravity target. Hook-based features need host support for hooks.

    Does Claude Code have built-in memory?

    Claude Code loads CLAUDE.md instruction files, which many of these tools write to. Search and capture come from add-ons like these.

    Are these safe to install?

    We have not run any of them. Our reviews are static code reads pinned to a commit. Read install scripts and hooks yourself first.