AI Memory Skillpack skill
Skills keep project memory in a docs/ai folder: short size-capped files, a one-page handoff and a verifier for the install
- 491 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The agent itself runs the installer: it edits the global `AGENTS.md` and `CLAUDE.md` and creates skill symlinks
- The `docs/ai/` files live in the repository and can reach a public git together with the memory contents
- The ban on writing secrets rests on rules and pattern-based sanitizing, so there is no full guarantee
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/ai-memory-skillpackDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The agent itself runs the installer: it edits the global `AGENTS.md` and `CLAUDE.md` and creates skill symlinks
- The `docs/ai/` files live in the repository and can reach a public git together with the memory contents
- The ban on writing secrets rests on rules and pattern-based sanitizing, so there is no full guarantee
This is third-party code. Review the repository files before installing.
What it does
AI Memory Skillpack turns the `docs/ai/` folder of a repository into project memory with hard size limits. It holds a project card up to 60 lines, architecture up to 250, a runbook up to 150, a gotchas list up to 300, ADR decisions up to 80 lines each and a `handoff.md` of up to 120 lines. `handoff.md` is replaced whole every time, and git keeps the history, so there are no separate change logs. Files fall into three reading tiers. The agent reads the project card and `handoff.md` whole at the start of a session, finds architecture, runbook and gotchas by keyword with `rg`, and opens `history/` only on a direct request. Before replacing `handoff.md` the skill rereads its header and takes a `flock` lock so two agents do not overwrite each other. Every edit is signed with its author in the `harness/model` format. The pack has three skills. `ai-project-memory` always installs, while `llm-wiki` and `project-mastery-coach` need a local wiki. The `INSTALL.md` installer is written for the agent: it adapts paths, merges the global `AGENTS.md` and `CLAUDE.md` rules and writes an install receipt. `VERIFY.md` checks the result by exit codes, and `SHA256SUMS` pins file integrity.
Who it is for. Maintainers and solo developers on Codex CLI, sometimes with Claude Code, who want the next session to know the project without rereading a mountain of notes.
Good fit when
- A second and third session on one repository spend tens of thousands of tokens on yesterday's notes
- Work passes between Codex and Claude Code: both read one `handoff.md` and sign their edits
- Secret values must not reach memory: `SYSTEM.md` forbids writing them
- Installed memory is upgraded to a new pack version and your own edits must survive: the receipt hash comparison leaves them alone
- The install result is checked by script: `VERIFY.md` and `sha256sum -c SHA256SUMS` give exit codes
Not a fit when
- You need a memory server with semantic search: this pack is files and `rg`, and the catalog lists memory servers such as Basic Memory
- A Windows machine without WSL: the installer needs Linux or WSL, bash 4.4 or newer, `git`, `rg` and `sha256sum`
- Memory must be in Russian: the skill tells the agent to write prose in Chinese by default, while commands and identifiers stay in English
- The project lives outside git: the install runs in a degraded mode and the commit rules do not apply
Example request
Set up docs/ai for this repository and record in the handoff where we stopped and which tests still failLimitations
- The pack's primary agent is Codex CLI, while Claude Code is supported as a second client through symlinks to the same skill files.
- The skills carry example paths and author strings (`claude-code/fable-5`, `codex/gpt-5.6-sol-pro`), and the installer adapts them to your machine.
- On macOS you need GNU `sha256sum`, `flock`, `readlink -f` and a recent bash, otherwise the installer stops.
- The wiki scripts are not in this repository, so `llm-wiki` and `project-mastery-coach` stay off without a wiki of your own.
How to disable. The install is reversible and the rollback is described in `INSTALL.md`, while the receipt with hashes shows what the pack installed. By hand, remove the memory sections from `AGENTS.md` and `CLAUDE.md`, the skill symlinks of the second client and the pack's skill directories. Delete `docs/ai/` folders one project at a time.
FAQ
How do I install AI Memory Skillpack?
Clone the repository, run `sha256sum -c SHA256SUMS` in it and tell Codex: "Read INSTALL.md in this pack and follow it to install the memory system on this machine." There is no GUI, because the installer is a program for the agent.
Does the pack work in Claude Code?
Yes, as a second client. The skill files live with the primary agent, Claude Code gets symlinks to them, and its memory rules go into `CLAUDE.md` separately because the clients' defaults differ.
What happens if I edited memory files by hand?
An upgrade compares file hashes from the receipt. Files you changed stay as they are, and the installer reports the mismatches.
Do I need a local wiki?
No. Without a wiki the installer skips `llm-wiki` and `project-mastery-coach` and installs only `ai-project-memory`.
Where does the agent write project memory?
Into the `docs/ai/` folder of the repository: `project-card.md`, `architecture.md`, `runbook.md`, `handoff.md`, `gotchas.md` and the `diagrams/`, `decisions/` and `history/` folders.
Discussion
No questions or reviews yet.
Related
Skillsai-memory retains session knowledge in a Markdown wiki, exposes MCP retrieval and passes context between agents and machines.
Long-term memory for AI agents: Markdown files as the source of truth, a SQLite index as a cache, shared by Claude Code, Codex and others
Persistent project memory in Markdown: decisions, requirements and constraints stay in the repo for each new session.
A plugin for Claude Code and Codex: a shared task wall, project memory, reports and a dashboard that shows agents' work across sessions
Agentic Substrate
Agentic Substrate (claude-user-memory)
A short reversible rule set for an agent: research the unclear, plan proportionally, verify the outcome and keep checked knowledge
A portable .agent folder with memory, skills and protocols that plugs into many coding agents and keeps its knowledge when you switch tools
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| AI Memory Skillpack (this card) | no | 491 | ||
| Basic Memory | no | not required | 4,495 | a local server with MCP, Markdown notes and semantic search, whereas this pack keeps memory as files in `docs/ai/` |
| Claude-Mem | no | 113,534 | a plugin that saves observations from sessions and feeds them into new ones, whereas this pack caps the size of memory files |
README in short
The front page presents memory as three reading tiers and contrasts it with ordinary notes that only grow. Then come the pack contents, a four-layer architecture, security and quality requirements, design tenets and a table of skills. Installation is covered for three cases (fresh, upgrade, migrate), along with system requirements and the repository layout.
README badge
Original SKILL.md
--- name: ai-project-memory description: Maintain bounded, durable AI project memory in a repository's docs/ai/ pack. Use when Codex or Claude Code needs to create, read, update, or audit project memory (project-card, architecture, diagrams, runbook, handoff, gotchas, ADRs), install AGENTS.md memory rules, or sync a lightweight central LLM Wiki project entity for user-owned projects. --- # AI Project Memory ## Core Rule 每个仓库的 `docs/ai/` 是该项目记忆的唯一事实源;**git 提交是唯一账本**(change-log 类文件已于 2026-07-28 停用,历史封存于 `docs/ai/history/`,不得重建)。 ```text docs/ai/ ├── project-card.md # 卡片:≤60 行,替换 ├── architecture.md # 架构入口:≤250 行,替换/局部更新 ├── diagrams/ # README ≤60 行 + *.mmd(一图一文件) ├── runbook.md # 操作手册:≤150 行,替换/局部更新 ├── handoff.md # 单快照:≤120 行,替换 ├── gotchas.md # 耐久陷阱:≤300 行,追加 + 定期精选 ├── decisions/ADR-*.md # 单篇 ≤80 行,新篇追加,旧篇冷存 └── history/ # 封存区,默认不读 ``` 记忆散文默认中文;命令、路径、代码标识符、config key、版本号保持英文。不写 secret 值。预算上限见上方结构图;仓库 `AGENTS.md` 只可声明更严(更小)的上限,不得放宽——VERIFY 硬门按固定常量执行,不识别放宽。 ## 读取纪律(检索先行) - **HOT(开机整读,口径 = 字节÷3)**:仓库 `AGENTS.md`、`CLAUDE.md`、`project-card.md`、`handoff.md`——四文件自身目标 ≤7k tokens;四文件 + 当前触发的 skill 正文合计 ≤10k。 - **WARM(按需定向,禁止整读)**:`architecture.md`、`runbook.md`、`gotchas.md`、`diagrams/README.md`、`decisions/`(现行 ADR)——用任务关键词(路径、symbol、config key、报错信息)`rg` 定位小节后只读该节。 - **COLD(默认不读)**:`history/`、`reports/`、`screenshots/`、`learning/`、旧 ADR(Status: superseded/deprecated)。仅任务明确要求追溯时才进,读到的内容必须与当前代码交叉核对后才能引用。 - **非分层新条目**:向 `docs/ai/` 一级新增任何文件/目录,必须同时在仓库 `AGENTS.md` 声明其层级;未声明即 COLD 且属违规(VERIFY 白名单断言会拦)。 ## 写入纪律 1. **谁干活谁写**:Codex 与 Claude Code 均可写,完成实质任务的一方负责更新。 2. **强制署名**,actor 只有两个拼法:`claude-code/fable-5`、`codex/gpt-5.6-sol-pro`。 - `handoff.md` 头部两行,固定列表项形式:`- updated: <ISO8601>`、`- updated_by: <actor>`(守卫与工具用 `grep -m1 '^- updated: '` 读取基线) - 追加条目(gotchas / ADR)末行:`— by <actor> · YYYY-MM-DD` - git 提交:Co-Authored-By trailer 对应同一 actor 3. **状态用替换**:`handoff.md` 永远是单快照 replace-in-place,不追加历史;追加语义仅限 gotchas 条目与新 ADR。