Search skills/

memU Claude skill

Memory layer that mines agent session logs into a shared markdown wiki and re-injects it across Codex, Claude Code, Cursor and others, locally or via memU Cloud.

NevaMind-AI/memU 1 skill Memory & contextSkill & agent tooling
Our verdict Recommended with caveats

Capable and well tested, but it reads your agent transcripts, edits instruction files, offers cloud storage and starts from a remote SKILL.md, so review the guide first.

4/5 Depth expertise shipped ! Security Broad access 0d Maintenance Updated today 15K Adoption stars ? License no clear license

The author changed the skill files after our review (noticed Oct 29, 2026). The review describes the earlier version; we recheck changed skills weekly, the most significant changes first.

What memU does

memU connects to your coding agents through two seams. The record seam is a scheduled task that slices new session logs into job files, has the agent itself distill them into memory and skill markdown, and commits the result. The inject seam patches a standing instruction into the host's instruction file, such as ~/.codex/AGENTS.md or a per-project AGENTS.md for Cursor, so the agent retrieves from memory in later sessions. The repo's SKILL.md is only an installer router: it tells an agent to identify its host, print the packaged guide and follow it to wire up or remove memU. Adapters exist for Codex, Claude Code, Cursor, OpenClaw, Hermes Agent and others, each reading that host's session transcripts. Memory can live locally in SQLite under ~/.memu or in memU Cloud, selected with MEMU_MEMORY_MODE, with an API key from memu.so. The README's quick start tells you to message your agent to read a SKILL.md hosted on memu.pro and follow it. The project has about 60 scripts, 28 reference files, 48 test files and CI, and the core memory logic is described as roughly 500 lines.

Good for

  • Sharing lessons across several coding agents
  • Turning repeated agent work into reusable skills automatically
  • Users comfortable with a scheduled background task reading their session logs

Not for

  • Sensitive codebases where transcripts must not leave the machine
  • People who want a passive, read-only memory tool

What's inside depth 4/5

A short 8 KB installer skill backed by about 28 reference files, 60 scripts, 48 test files and CI in a 288-file repo.

Skills
1
Reference docs
28
Scripts
60
Tests / evals
48
Templates
24
Avg SKILL.md
7.9 KB

Counted from the repository file tree. How we score depth.

Dependencies and costs

  • memU Cloud account and API key (optional), or a local store with your own embedding provider
  • A scheduler on the host for the record task
  • Python or npm packages that ship the memu host binaries
  • Telemetry: In cloud mode, distilled memory is hosted by memU Cloud; local mode keeps it in a SQLite file under ~/.memu. The dossier does not describe any other telemetry or an opt-out.

Security review

Broad access

This skill asks for more access than most, or sends data to another service. That is usually a design choice rather than bad intent, but you should know before you install it.

  • Loads instructions from a server at run timeThe self-evolve job prompts and the retrieval skill are fetched from https://memu.pro/sdk/instructions/ (refreshed daily, cached) and then followed by a headless agent or written into your instruction file; MEMU_TEMPLATE_BASE_URL='' turns it off.src/memu/hosts/templates.py:53
  • Installs hooks that run automaticallyRegisters UserPromptSubmit hooks (retrieval and background memorization) in Codex's hooks.json and writes their trusted_hash into Codex config itself, so Codex's own hook-trust step is satisfied programmatically.src/memu/hosts/codex/prompt_hook_install.py:182
  • Turns off permission prompts or safety checksThe background memorization worker runs codex exec --dangerously-bypass-approvals-and-sandbox, and SKILL.md asks the Codex user to enable Full access / the same flag before install.src/memu/hosts/codex/cli.py:58
  • Turns off permission prompts or safety checksTells the agent to install in one pass, not stop for confirmation, and take defaults for every choice (hourly schedule, cron, official installers).SKILL.md:105

memU is a long-term memory system for coding agents: install patches your global instruction file, adds hooks or an hourly scheduled job, and mines your past session transcripts into memory files, stored locally (SQLite under ~/.memu) or, if you provide a key, in memU Cloud. The job prompts and retrieval skill are fetched from memu.pro at runtime and followed by a headless agent that, for Codex, runs with the sandbox bypassed, and the Codex hooks are auto-trusted by the installer. Event-count telemetry is on by default and can be disabled with MEMU_TELEMETRY=0 or DO_NOT_TRACK=1; only part of the code was read.

Worth knowing

  • Sends usage data by default (can be turned off)Reports lifecycle event counts to https://api.memu.so/api/memu/analytics/events by default; MEMU_TELEMETRY=0, DO_NOT_TRACK=1 or an empty MEMU_EVENTS_BASE_URL disables it; file says no query text, memory or paths are sent.src/memu/events.py:78
  • Runs remote code at an unpinned versionThe npm launcher runs uvx --from memu-cli memu (or pipx run) with no version pin, and SKILL.md installs with pip install --upgrade memu-cli.npm/bin/memu.js:33
  • Installs with curl | shDocs and host guides tell you to run curl ... | sh for the uv installer (astral.sh) and for the Claude Code, Codex and Cursor installers when missing.INSTALL-LATEST.md:48
  • Changes files outside your projectInstall patches global instruction files (~/.claude/CLAUDE.md, ~/.codex/AGENTS.md), adds a skill under the host's skills directory, edits ~/.claude/settings.json permissions, registers an hourly crontab/launchd/Task Scheduler job, writes ~/.memu/config.env with API keys and may persist an Anthropic key in the crontab or user environment.src/memu/hosts/instruction.py:237
  • Installs packages or toolsInstalls the memu-cli Python package (pip or uv tool) and its host-adapter binaries onto PATH.SKILL.md:34
  • Makes network requestsIn cloud mode sends distilled memory to and retrieves from https://api.memu.so; also calls memu.pro for templates/docs and any configured embedding provider (OpenAI, Jina, Voyage, OpenRouter, Doubao).src/memu/cloud.py:13
  • Needs an API key for its own serviceMEMU_CLOUD_API_KEY is sent only to the memU cloud API; embedding-provider keys go only to their own providers.src/memu/cloud.py:99
  • Relies on a CLI or package we could not reviewDrives the host CLIs codex, claude -p and cursor headlessly on a schedule; their code is not in the repo.src/memu/hosts/codex/cli.py:50

What we checked

The repository (288 files) on Oct 29, 2026. Not a full review: we read 11 of 187 code files; it relies on code outside the repository: Most of the 187 code files (database repositories, tests, per-host session readers, host_cli.py, instruction.py, windows scheduler) were not read in full; PyPI package memu-cli as published (may differ from this source); Server-side templates and docs served from https://memu.pro/sdk/instructions/ and the memU cloud API.

NVIDIA SkillSpector static scan: 228 pattern hits (Rogue Agent 98, Data Exfiltration 35, Dangerous Code Execution 31, Agent Snooping 15, Excessive Agency 12, Supply Chain 9). The CRITICAL tainted-flow hit is templates.py fetching templates from the MEMU_TEMPLATE_BASE_URL setting (a real remote-instructions path, flagged). Credential and '.netrc' hits are config handling and httpx proxy behavior; the 'backdoor_persistence' YARA hit is the uninstall guide telling the agent to list crontab/LaunchAgents to remove memU's own entries; 'self-evolve' is the product's own memory-mining job name.

A review, not a certification. Skills change after we check them — read the files before installing.

License

No recognized license (GitHub reports NOASSERTION); read the LICENSE file before commercial use.

Can it run unattended?

Not really. It runs in the background to record memory, but it is a memory layer and does not drive tasks or loops.

Alternatives to memU

memsearch

zilliztech/memsearch

Cross-agent persistent memory for Claude Code, Codex, OpenClaw, OpenCode and DeepSeek Harness: Markdown files as the source of truth, with a Milvus search index.

With caveats Broad access Depth 3/5

Obsidian Second Brain

eugeniughelbur/obsidian-second-brain

Persistent, self-rewriting memory in your Obsidian vault: 48 commands, a shared cross-agent logbook (Crew), scheduled maintenance and a research toolkit.

With caveats Check first Depth 5/5

Continuous Claude

parcadei/Continuous-Claude-v3

A personal Claude Code infrastructure: ~160 skills, 32 agents and 30 hooks that inject skill hints, isolate subagent context to files, and fight compaction.

With caveats Broad access Depth 3/5

Context Mode

mksglu/context-mode

MCP server plus hooks that keep bulky tool output out of the context window, index it in local SQLite, and restore session state after compaction.

With caveats Broad access Depth 3/5

FAQ

What is memU?

memU is a personal memory system for coding agents. It distills your session history into a markdown wiki and reinjects it into later sessions across agents and devices.

How do I install memU?

The README's quick start has you message your agent to read https://memu.pro/SKILL.md and follow it, with a memU API key for cloud mode. A local install, which keeps memory in SQLite, has its own message in the README.

Does memU send my data to the cloud?

Only if you pick memU Cloud, where memory is hosted by the service. In local mode memory stays in a SQLite file under ~/.memu, per the README.

memU vs context-mode?

memU keeps long-term memory distilled from past sessions across agents. context-mode focuses on keeping bulky tool output out of the current context window and restoring state after compaction.

Sources: GitHub repository (stars, license, file tree, last push Oct 29, 2026). Review based on the README and SKILL.md files as published.