Dynamic Memory
lleoliang/dynamic-memory
Memória persistente e dinâmica para Claude como colaborador de longo prazo
Ver no GitHubUma skill de memória dinâmica que transforma Claude em um colaborador de longo prazo, capaz de decodificar atalhos, lembrar preferências e contexto do usuário. Organiza informações em categorias (pessoas, projetos, glossário, locais) com cache quente para acesso rápido e armazenamento profundo para dados completos.
Instalação
npx convez-ai lleoliang/dynamic-memory1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Compartilhar contexto entre múltiplas sessões Claude para tarefas diferentes
- Manter memória coletiva de equipe em repositórios ou diretórios compartilhados
- Armazenar contexto específico do projeto dentro do repositório de código
Veredito

Skill prática e bem pensada para quem precisa de continuidade real entre sessões Claude. O sistema de cache quente + armazenamento profundo é eficiente; o suporte a categorias dinâmicas evita engessamento. Requer disciplina do usuário para manter memória útil e não poluída.
README
A tradução para o português deste README ainda está na fila.
dynamic-memory
A memory skill for productivity that makes Claude a long-term collaborator—someone who understands your language and remembers everything you've told it.
What it does
Without memory, shorthand is meaningless. With dynamic-memory, Claude decodes it instantly:
User: "ping j about horizon before standup"
↓ Claude decodes
"Message Jamie Martinez (Engineering Manager) about the Horizon project
(mobile app redesign) before the 9am daily standup in #eng"
It also remembers your preferences, stores contact profiles, tracks projects, and can create new memory categories on the fly—all automatically indexed for fast lookup.
Useful Applications
- Shared memory across sessions — you can connect multiple Claude sessions to the same
memory/folder and they all read from and write to the same memory. This is useful if you run separate sessions for different tasks (e.g. coding, scheduling, research) but want them to share the same understanding of your people, projects, terminology, etc. - Team memory — put
memory/in a shared directory or repo so multiple people on a team contribute to and benefit from the same knowledge base. - Per-project memory — keep a
memory/folder inside each project repo for project-specific context that travels with the codebase.
Installation
Add SKILL.md to your Claude project or skills directory.
On first use, Claude will initialize the memory/ directory automatically. If a memory/ directory already exists, Claude will detect it and ask whether to keep the existing memory or start fresh. It will never overwrite existing memory without confirmation.
Just say "set up memory" or start working — Claude will build it as you go.
How it works
Two tiers:
-
Hot cache (
memory/cache.md) — loaded once at session start. Covers ~90% of daily needs without reading any other file. Top ~30 people, ~30 common terms, active projects, frequent locations. -
Deep storage (
memory/) — full profiles, complete glossary, project details, and any custom category. Looked up only when the hot cache misses.
memory/
cache.md ← Hot cache (~100 lines, read at session start)
preferences.md ← User preferences — flat file, read at session start
index.md ← Registry of all categories
glossary.md ← Full decoder ring (terms, acronyms, nicknames)
people/ ← One file per person
projects/ ← One file per project
locations/ ← Physical places, drive times
[any-category]/ ← Dynamically created, registered in index.md
Default categories
| Category | What it stores |
|---|---|
people/ | Contact profiles, nicknames, roles |
projects/ | Project details, codenames, status |
glossary.md | Terms, acronyms, shorthand (flat file) |
locations/ | Addresses, drive times, hours, notes |
Dynamic categories
The key feature: categories aren't hardcoded. When something doesn't fit the defaults, Claude proposes a new category:
"This doesn't fit any existing category. I'd suggest creating a
job-applications/category for tracking job applications. Should I go ahead?"
On confirmation, it creates the directory, registers it in index.md, adds a hot cache section, and creates the first entry.
Preferences
Just say what you prefer — no special syntax:
"I prefer short answers"
"always block drive time before appointments"
Preferences are saved to memory/preferences.md automatically. Whenever a preference influences a decision, Claude names it explicitly so you can catch and correct stale preferences.
Portability
All memory lives in memory/ — no CLAUDE.md edits required. Move the directory to any project and memory comes with it.
Interesting Considerations
One might notice that we are storing preferences in a flat file (i.e. one markdown file) instead of the directory format (e.g. people/, locations/). This was a natural consequence of realizing that preferences lack the same "categorizable" trait that other things like locations may have. The directory format works well when entries have a natural identity—a person has a name, a location has an address, a project has a title. Each entry is a discrete, nameable thing you can look up by key (e.g people/jamie-martinez.md).
I found that preferences don't work that way. They're behavioral patterns that apply globally. They are not things you look up, but things Claude reads in full and holds in context at all times. There's no meaningful "key" to index a preference under, and splitting them into individual files (e.g. prefer-async.md, block-drive-time.md) would just create an arbitrary structure with no lookup benefit.
Categorizing preferences (e.g. style, workflow, domain) was considered but rejected for the same reason: attempting to force structure upon preferences would lead to unncessary overhead. Plus, Claude already understands the context of a preference naturally without needing a folder to signal it.
Contributing
This repository is open to contributions and new ideas. If you have a suggestion, improvement, or a new use case in mind, feel free to open a pull request — I'll review it.
Acknowledgments
Inspired by the memory-management component of the Productivity plugin by Anthropic—with a key difference: that system only stores a few fixed categories while dynamic-memory supports unlimited categories. Another difference in this version is that the hot-cache is stored in memory/cache.md instead of CLAUDE.md; this way the dynamic-memory system is completely independent and won't compete with other components for CLAUDE.md