Claude Code Session Memory: Use Productivity and CLAUDE.md Plugins to Preserve Project Context

In a first-person piece published October 21, 2026, XDA Developers writer Nolen Jonker describes using two Anthropic-made plugins to cut down that catch-up time. One is Productivity and the other is claude-md-management. The headline says Claude "picks up where I left off," but that oversells it a little. Neither plugin restores a full conversation. What they do is help you write selected project state and lasting rules into plain files, and Claude reads those files when the next session starts. That's still useful. You just need to know which parts are documented behaviour and which parts are one writer's experience.

What Claude Code already carries between sessions​

It helps to know what Claude Code does on its own first. Anthropic's official memory documentation describes two mechanisms:

• CLAUDE.md files are written by you and contain instructions and rules. They load into every session.
• Auto memory is written by Claude itself. It holds learnings and patterns, and the first 200 lines or 25KB load into every session.

Jonker's complaint is that auto memory decides what to keep without asking you, and Anthropic's docs confirm this. Claude doesn't save something every session. It decides what's worth remembering based on whether the information would be useful in a future conversation. The docs also say Claude skips anything it can derive from the codebase, such as architecture, file paths, or debugging fixes. It also skips anything your CLAUDE.md files already say.

The notes also aren't stored in your project folder. Several independent guides, Archcore's among them, report that Auto Memory defaults to a directory outside your repository; autoMemoryDirectory can change that location. Vectorize's guide gives the default path as ~/.claude/projects/<project>/memory/. On Windows, that's the .claude folder in your user profile.

You might think CLAUDE.md should cover the gap, but Anthropic doesn't recommend treating it as a running diary. The official guidance says to target under 200 lines per CLAUDE.md file. Longer files consume more context and reduce adherence. Notes like "here's where I left the project yesterday" don't really belong in a file that loads into every session.

Section summary: Claude Code's built-in memory is either curated by you (CLAUDE.md) or chosen by Claude (auto memory). Neither is designed to track day-to-day project state, and that's the gap these plugins fill.

Plugin one: Productivity, a task tracker that remembers your shorthand​

Productivity lives in Anthropic's knowledge-work-plugins repository. Its README says it is primarily designed for Cowork, Anthropic's agentic desktop app, but that it also works in Claude Code. Jonker says that suited their workload, which is mostly computer admin rather than coding.

According to the README, the plugin does three things:

• Task management through a Markdown file, TASKS.md. Claude reads it, updates it, flags stale items and can sync with external tools.
• Workplace memory in two tiers: CLAUDE.md holds working memory and a memory/ directory holds deeper notes, so Claude learns your people, projects and terminology.
• A visual dashboard, a local HTML file that shows your tasks as a board and what Claude currently knows. Edits made on the board or in the file stay in sync.

Installing it​

The repository lists these terminal commands:

claude plugin marketplace add anthropics/knowledge-work-plugins
claude plugin install productivity@knowledge-work-plugins

Jonker notes that plugins can also be installed from the desktop app. Check the commands against your installed CLI version, because plugin tooling changes quickly.

Using it​

The README documents three commands:

The command names differ between sources. The README uses /start and /update, while Jonker types /productivity:start and /productivity:update. The prefixed form is most likely the plugin namespace appearing in Claude Code. If one form doesn't work, try the other and look at what your slash-command menu shows.

The README's own example shows /start creating TASKS.md, CLAUDE.md, a memory/ directory and dashboard.html, then asking about your role and priorities. Jonker says their file came out split into Active, Waiting On, Someday and Done.

The rest of Jonker's account is personal experience, not documented behaviour:

• They pasted in an existing list. The plugin sorted it and turned a vague "before Friday" into an actual date.
• Updating a task meant saying something like "that one's waiting on feedback," and Claude moved it to the right column.
• When it met their abbreviations, it asked what each one meant and which project it belonged to. It saved the answers to CLAUDE.md and the memory folder.
• In a new session, the update command now goes through every task, including how many days are left on anything with a due date.

The README also lists optional connectors for Slack, Microsoft 365 email, calendar and Office, Notion, and project trackers including Asana, Linear, Atlassian, monday.com and ClickUp. It also says you can manage tasks and memory by hand without any of them. External sync only happens if you set up those connections.

Section summary: Productivity keeps your current work state in files you can read yourself (TASKS.md, memory/, a dashboard), so you don't have to rely on Claude's own memory choices.

Plugin two: claude-md-management, turning corrections into rules​

The second plugin, from Anthropic's claude-plugins-official repository, is very small. The README describes two parts that work together:

• /revise-claude-md (a command) captures what was learned in the current session. It's meant for the end of a session, when the session has revealed something missing from CLAUDE.md.
• claude-md-improver (a skill) checks CLAUDE.md files against the current state of the codebase. It's meant for occasional maintenance, and you trigger it with plain requests such as "audit my CLAUDE.md files."

Jonker's routine works like this. During a session, they tell Claude something worth keeping, such as a naming rule for new folders. Before closing, they run /revise-claude-md. Claude reviews the session and drafts additions as a diff, and nothing is written until they approve it. Compare that with auto memory, which reports something like "saved 2 memories" without asking first.

When Jonker ran the improver, it scored their CLAUDE.md at 79 out of 100. It also flagged details they hadn't written down, such as folder names with spaces that can break commands. That's one person's result, not a benchmark.

The file-growth problem​

Jonker also points out a real limitation: /revise-claude-md suggests new lines but never removes old ones, so CLAUDE.md keeps growing. Other users have noticed the same thing. In a June 14, 2026 feature request on Anthropic's claude-code GitHub repository, a user described the command as following a "one-off, additive model" that appends lines without pruning or consolidating. They also said it ignores skill directories and sends everything to CLAUDE.md. GitHub's automation closed the issue as "not planned" on July 27 after it went inactive. That isn't a ruling either way, but it means no fix is publicly committed.

That matters because of Anthropic's own guidance. Third-party guides such as Vectorize's point out that CLAUDE.md files are loaded in full regardless of length (though shorter files produce better adherence). Each approved line costs context in every future session. Jonker's fix is to run the improver every so often to trim the file. If the improver can't trim enough, the Stele guide suggests that when a file grows, move the parts that only matter in one area into a path-scoped rule, and move multi-step procedures into a skill, which loads only when it is needed.

Section summary: claude-md-management makes you approve updates to CLAUDE.md, which beats letting Claude pick what to save. The file only grows, though, so you'll need to prune it yourself.

A practical end-of-session routine​

Here's a workflow built from the documented features and Jonker's account:

1. Set up once per project folder. Install Productivity, run its start command, and paste in your current to-do list.
2. Answer its questions. Explaining your abbreviations and project names is what makes later sessions useful.
3. Update tasks in plain language as work changes: "X is blocked on legal," "Y is done."
4. Start each new session with the update command to check where things stand.
5. Before closing a session, run /revise-claude-md only if the session turned up a lasting rule. Read the diff and turn down anything that's really temporary state.
6. Audit regularly with claude-md-improver, and keep CLAUDE.md under roughly 200 lines.
7. Check what's loading. Vectorize notes that /memory is the management entry point. From there you can see which files are loaded, switch auto memory on or off, and edit memory files directly.

The key idea is to keep standing instructions separate from changing work state. Naming rules and build commands belong in CLAUDE.md. Blockers, deadlines and next steps belong in TASKS.md. If you mix the two, you end up with a bloated file that Claude follows less reliably.

One general caution: an audit skill that checks documented commands is only as trustworthy as the files it reads. Be careful about approving commands taken from CLAUDE.md files in cloned repositories or dependency folders you didn't write.

The bottom line​

Neither plugin gives Claude perfect recall, and nothing Jonker describes claims it does. The persistence comes from keeping a few files up to date, and the plugins handle the tedious parts: writing, sorting, asking questions and asking for approval. Jonker says the biggest change was how they end sessions. Instead of closing the window, they spend a moment telling Claude what got done and approving new rules.

Anyone who has lost a morning re-explaining a project to a fresh session will see the appeal. You still have to review the diffs and trim CLAUDE.md now and then, but that's less work than starting from scratch every session.