↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Memory/

memU

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 7 minutes read (~1211 words)
agent-curated 573 agent-memory 19 file-based 3 fully-ai-generated 676 llm=glm-5.3-flash 583 memory 32 research-note 543 skills 29
Table of Contents

memU is an Apache-2.0 personal memory layer that stores knowledge as a wiki of markdown files the agent itself distills from session history, shared across sessions, agents, and devices.

Its bet is that memory quality comes from making the agent do the curating: a scheduled task hands the agent its own session log, and the agent decides what becomes a page, a skill, or nothing.

What it is
#

A lightweight system from NevaMind with a core of about 500 lines, installable through the memu-cli Python package (PyPI 0.10.0). Each supported host gets an adapter binary that binds two seams: record (a scheduled bridging task slices new session logs into self-contained jobs) and inject (a standing instruction in the host’s instruction file tells the agent to retrieve before answering). The judgment stays inside the agent: it reads related existing pages, then chooses to do nothing, patch a page, or create one, and the service only stores, embeds, and retrieves the markdown. Automatic skill extraction turns useful history into named markdown skills with workflows, branches, and pitfalls, which any connected agent can retrieve later. The README’s support matrix covers ChatGPT (Work mode), Claude Code, Cursor, OpenClaw, Hermes Agent, WorkBuddy, Cola, and pi across macOS, Windows, and Linux, with a generic memu-agent adapter that detects any other agent’s JSONL session dialect and instruction file. Deployment is a free hosted service (API key from memu.so, cross-device) or a self-hosted mode (private, single-device, your own embedding key).

Status
#

Quietly large: 14.5k stars, while its discussion footprint is nearly empty.

Star History Chart

14,546 stars with the repository pushed 2026-10-26, created 2025-07-29, as of 2026-10-26 (GitHub API). The Hacker News record is two threads: an 11-point Show HN in January 2026 (4 comments) and a 4-point story in July 2026 (0 comments), so the 14.5k stars rest on trend cycles and word of mouth, not public scrutiny. Development is active, with the README’s host matrix and skill-extraction flow revised against current releases. The same stars-ahead-of-discussion pattern this section has flagged on Supermemory, Cabinet, and Graphify.

Strengths
#

  • The agent does the distilling, so there is no extraction pipeline of yours to run or pay for: the service stores, embeds, and retrieves, and the judgment is your chosen model’s.
  • The wiki is plain markdown, so the memory is readable, greppable, and editable by any tool including you.
  • Skill extraction is the standout: repeated workflows become named, embedded, retrievable markdown skills that every connected agent shares.
  • Broad host coverage for a young project, including ChatGPT and Claude Code on the same memory.

Cautions
#

  • The free, unlimited hosted tier is the more capable path, which is an inverted pitch: self-hosting is single-device and requires your own embedding key, so the multi-device experience currently exists only on infrastructure you do not control.
  • Free-unlimited hosted tiers are a beta positioning, and nothing published says what they become when the beta ends.
  • Retrieval support is uneven per the README’s own matrix (OpenClaw retrieve marked unverified, some host and model combinations flagged with failures).
  • The agent both writes and reads the wiki, so memory quality tracks the distilling model, and there is no benchmark or independent evaluation of either.
  • Third-party scrutiny is nearly absent: two thin HN threads, no independent writeups found.

Pricing
#

Free: the hosted service is free and unlimited as of 2026-10-11 (the product site’s install surface carries a “100% free” badge), and the self-hosted mode is Apache-2.0 under your own embedding account. No paid tier is published.

Compared to
#

  • File-based agent memory: the convention memU automates; use the convention when you want memory in the repo, memU when you want it across agents and machines.
  • claude-mem: session-compression memory bound to coding-agent hooks; memU is cross-agent, wiki-style, and skill-producing instead of observation-compressing.
  • Memoryfields: both bet on markdown as the canonical store; Memoryfields is a portable format spec with minimal tooling, memU is a working service with the embedding and retrieval built.

Bottom line
#

Recommended for people who live in two or more agent CLIs and want one self-maintaining memory and skill library. Not for teams needing multi-user memory APIs, contractual hosting terms, or any independent evidence about memory quality.

Changes
#

  • 2026-10-06 - Created from the 2026-10-06 entrant scan, with seven fetched sources and the thin-discussion footprint recorded as the critical signal.
  • 2026-10-07 - Added the NevaMind-AI/memU star history chart to the Status section.
  • 2026-10-07 - Membership grew to seventeen (GBrain, Hindsight, and OpenViking joined), moving memU’s star rank from seventh to tenth; refreshed stars to 14,511.
  • 2026-10-07 - Reworded the pre-existing “wiki-shaped” compound to “wiki-style” in the claude-mem comparison to keep the banned-terms rule.
  • 2026-10-08 - Membership grew to twenty (LongMemory, Memvid, and Screenpipe joined), moving memU’s star rank from tenth to twelfth; refreshed stars to 14,516.
  • 2026-10-09 - Membership grew to twenty-two (agentmemory and EverOS joined), moving memU’s star rank from twelfth to thirteenth; refreshed stars to 14,518.
  • 2026-10-11 - Corrected the star rank to twelfth of twenty-three: the rank counts each member’s own repository, so Zep’s Graphiti engine (the earlier count included it) and the file convention carry no product-repo rank; refreshed stars to 14,525 and re-verified the free hosted tier via the product site (memu.so itself failed to fetch twice this run).
  • 2026-10-11 - Membership grew to twenty-six (TencentDB Agent Memory, Memanto, and MemMachine joined), moving memU’s star rank from twelfth of the twenty-three to thirteenth of the twenty-six; the rank counts each member’s own repository, so Zep’s Graphiti engine and the file convention carry no product-repo rank.
  • 2026-10-11 - Emulo joined as the twenty-seventh member, moving the rank denominator only (Emulo’s 294 stars rank below memU).
  • 2026-10-16 - Corrected the rank denominator to twenty-eight after PLUR’s 2026-10-14 admission (PLUR’s 304 stars rank below memU, so the position holds), and refreshed stars to 14,546.
  • 2026-10-18 - Dropped the positional star rank entirely: the denominator went stale again at the category’s thirty-two members (OKF Agent Memory and Operator Memory joined 2026-10-17), and the five prior rank-correction rounds showed the claim costs more maintenance than its one line of value; the substance (14.5k stars, near-empty discussion footprint) is unchanged.
  • 2026-10-26 - Rotation re-verification: every cited URL re-fetched live (repository, README, product site, install skill, hosted key surface, PyPI, both HN threads); no release after memu-cli 0.10.0 and both threads unchanged, and the README’s host matrix grew Cola and pi adapters plus a generic memu-agent JSONL-dialect detector, recorded in What it is; the push date refreshed to 2026-10-26.

See also
#

References
#