Rishi 发布了开源的 remembrandt,用 Skill 和 CLI 记录项目决策,并在后续编码会话中读回。记忆由本地 JSON 条目和生成的 Markdown 摘要保存。
remembrandt:把项目决策留给下一次会话
remembrandt 将 Agent Skill 与 Python CLI 配合使用,把项目决策、踩坑记录和已完成的工作存成本地 JSON,再整理为可阅读的 MEMORY.md。
来源信息
- 仓库
- rishibanota/remembrandt
- 最近来源活动
- 2026年8月8日 07:56
- 检测到的 SKILL.md 语言
- 英语
- 星标
- 1
- 分支
- 1
用途
来源要求 Agent 在会话开始时读取 .remembrandt/MEMORY.md,工作中记录重要决策和问题,结束时重新生成摘要。记录可归入 decision、changelog、gotcha、note 或 context 等类别。
必要前提
README 要求 Python 3.9 及以上,CLI 使用标准库。需要同时安装 CLI 和 Skill;仅复制 SKILL.md 不会让 remembrandt 命令出现在 PATH 中。源文件还提供了在项目根目录用 python -m agent_memory 调用的方式。
操作说明
作者 README 演示了记录一条决策并生成会话摘要:
remembrandt remember "Use uv for dependency management" --category decision --tags uv
remembrandt recall
remembrandt resume
最后一步写出 .remembrandt/MEMORY.md。
限制
记忆保存在项目文件中,可提交给团队共享,也可放入 gitignore 留在本地。现有流程需要实际记录条目,并读取生成的摘要;通过 git hooks 自动捕获和 MCP server 在 README 中属于后续计划。项目采用 MIT 许可证。
安装方式
默认使用会先检查来源的 Prompt;你也可以切换为直接命令,或下载本地副本。
检查来源文件
决定是否安装前,请先阅读 SKILL.md,以及 SkillsMP 当前展示的配套文件。
正在显示 SKILL.md
SKILL.md
来源说明 · 只读预览- name
- remembrandt
- description
- Persistent project memory for coding agents. Use this skill to never lose context across sessions. At the start of work, read the project's MEMORY.md to recover decisions, gotchas and the changelog. During work, capture important decisions, gotchas and progress so they survive session restarts. Call the `remembrandt` CLI to remember, recall, search and resume.
- metadata
- {"version":"0.1.0","license":"MIT","tags":["memory","agent","claude-code","cursor","codex","context"]}
# remembrandt — Persistent Agent Memory
You are the memory layer for this project. Your job is to make sure nothing
important the human or another agent decided is ever lost, even across
completely fresh sessions.
## When to do this
1. **On session start** — if a `.remembrandt/MEMORY.md` file exists, read it
*first* before doing any other work. It is the single source of truth for
project context: prior decisions, gotchas, and the changelog.
2. **Whenever a meaningful decision is made** — record it so future sessions
(or the human) know *why* the code is the way it is.
3. **Whenever a gotcha / pitfall is discovered** — record it so the same bug
is never reintroduced.
4. **Whenever a task is completed** — append to the changelog.
5. **At the end of a working session** — regenerate `MEMORY.md` with
`remembrandt resume` so the human can review and commit it.
## How to use the CLI
```bash
remembrandt remember "Use `uv` for dependency management" --category decision --tags uv
remembrandt remember "Never mutate the shared config object" --category gotcha
remembrandt recall # list everything, newest first
remembrandt recall decision # filter by category
remembrandt search uv # search all entries
remembrandt resume # write .remembrandt/MEMORY.md
remembrandt status # show stats
remembrandt forget 12 # delete entry #12
```
If the `remembrandt` executable is not installed, run the equivalent
`python -m agent_memory ...` commands from the project root.
## Recommended categories
- **decision** — architectural and code decisions and their rationale
- **changelog** — what changed and when (a lightweight changelog)
- **gotcha** — pitfalls, sharp edges and things to never do again
- **note** — general useful notes
- **context** — background, requirements, constraints about the project
Keep entries short, specific and self-contained. Prefer many small entries
over one giant blob. Add tags when it helps searching.
在 GitHub 查看