Memory Management
Two-tier working-memory guidance for shorthand, people, projects, and internal workplace language.
- Skill Road
- Memory Management
Categories
Memory Management is an individual skill from Anthropic’s official knowledge-work-plugins repository. Its direct primary source is https://github.com/anthropics/knowledge-work-plugins/tree/main/productivity/skills/memory-management. The skill is not a database, a standalone agent, or a promise of memory in every Claude environment. It describes a maintainable storage and retrieval method that helps Claude interpret workplace acronyms, nicknames, projects, roles, tools, and internal terminology. The instructions are most useful when short requests need to make sense to a team without requiring people to repeat the full context in every conversation.
Goal and architecture
According to the provider, the approach has two tiers. The CLAUDE.md file acts as a compact hot cache for information that is needed frequently. It can hold key contacts, common acronyms, active projects, and working preferences. The deeper knowledge base lives in a memory directory. A glossary.md file can collect the complete vocabulary, while people, projects, and context directories hold detailed profiles, project records, and company or team context. This separation is intended to keep frequent lookup fast while allowing the documentation to grow. Anthropic recommends keeping the fast layer deliberately small rather than copying every historical detail into the central file.
Tiered retrieval
For a request, the instructions say to check the working file first. When a nickname, acronym, or project name is found there, Claude can translate the shorthand into a more useful working context. If the item is missing, the next step is to search the full glossary. For an important action, the relevant person, project, or company profile can then provide deeper context. If the term remains unknown, Claude should ask the user instead of inventing a plausible meaning. This is a practical progressive-disclosure strategy: frequent information stays compact, while rare details are opened only when needed.
Content and maintenance
The source presents tables as a compact format for people, terms, and projects. A useful person record should include more than a name: it can capture role, team, alternate names, and relevant working context. Project records benefit from a codename, status, participants, and relationships to other initiatives. Frequently used information can be promoted into CLAUDE.md. Stale or rarely needed details should instead be demoted to deep storage. For initialization, Anthropic names the productivity start command, which can create task and memory structures. The exact availability of such commands depends on the Claude product and its plugin integration.
Boundaries, privacy, and trust
Memory Management does not create independent truth. An incorrectly maintained name, an outdated project, or an ambiguous acronym can produce a wrong interpretation. Review recipients, deadlines, permissions, and external actions before sending or executing anything consequential. Workplace profiles can contain personal, confidential, or security-sensitive information. Store only approved details, minimize context, define retention and access rules, and never put passwords, tokens, or other secrets in these files. A local Markdown knowledge base also does not automatically mean that a connected model runs locally; the selected client may transmit content to its model provider. The repository is licensed under Apache-2.0. This description was checked against the official source on September 9, 2026 and describes the method, not the correctness of any particular company documentation.
Installation instructions
What the Memory Management skill does
The Memory Management skill comes from Anthropic's official knowledge-work-plugins repository and is part of a collection of productivity skills for Claude. According to the provider, the skill turns Claude into a genuine workplace collaborator that understands internal shorthand, nicknames, and company jargon instead of requiring clarification on every request. A typical scenario: a user types something brief like "ask todd about the psr for oracle," and Claude needs to infer who Todd is, what PSR stands for, and which project is meant by Oracle. Without a memory system, such a message is meaningless to an AI model because the context only exists inside employees' heads.
The two-tier memory architecture
At the core of the skill is a two-tier storage architecture. The first tier is a file called CLAUDE.md in the current working directory, acting as a fast working cache: according to the documentation it holds roughly the top thirty people a user interacts with, the most common acronyms and terms, and active projects, kept compact in table form and capped at around fifty to eighty lines. This file is designed, per the skill description, to cover ninety percent of daily decoding needs so Claude does not need to run an expensive search for every small request. The second tier is a memory directory containing a full glossary plus subfolders for people profiles, project details, and general company context. If Claude cannot find a term in the fast cache, it first searches the glossary and only then, if necessary, the more detailed profile and project files.
The lookup flow in practice
The flow follows a clear order: first CLAUDE.md is checked, then if nothing is found the broader glossary at memory/glossary.md is searched, and only if that still comes up empty does Claude actively ask the user and remember the answer for future requests. This tiered structure keeps the fast layer lean while the deep-memory layer can grow essentially without limit, without slowing down everyday responsiveness. For teams working with many internal abbreviations, project codenames, and recurring people, this mechanism cuts down on having to re-explain context in every new chat session.
Step 1: Prerequisites and setup
To use the skill, it is loaded into the relevant Claude context as a skill plugin, typically through the plugin mechanism of the Claude environment where skills from official or community repositories can be installed. Afterward, a CLAUDE.md file is created in the project or working directory and populated with the most important people, terms, and projects following the table format suggested in the documentation. For deeper information, a memory subdirectory is set up containing a glossary.md file along with people, projects, and context folders. Maintaining these files is the user's own responsibility; the skill only defines the format and the lookup logic, it does not populate the content automatically.
Security, limitations, and best practices
Because CLAUDE.md and the memory files typically contain personal data such as names, roles, and internal business information, they should not accidentally end up in public repositories or shared contexts. It is advisable to keep them outside version-controlled, publicly accessible directories or to exclude them explicitly via a gitignore rule. Another security consideration is staleness: outdated entries about completed projects or former colleagues can lead to incorrect associations, so regular maintenance is worthwhile. According to the description, the skill itself is not directly user-invocable but works in the background as a contextual aid. Its limitation is that it is only as good as the data that has been maintained; it does not generate knowledge automatically but merely structures how existing knowledge is looked up. For individual users without a complex team context, the added value is correspondingly smaller than for teams juggling many internal abbreviations and recurring contacts.
- Provider
- Anthropic
- License
- Apache-2.0
- Last reviewed
- 09.09.2026