All video notes
Video notes · Claude Code memory

Claude Code's Hidden /dream Feature MASSIVELY Upgrades Memory

Video by Chase AI on YouTube · 9 min · uploaded March 2026

Notes and analysis by BrainTube, written from the video. Not affiliated with Chase AI. Watch on YouTube

In short

Chase explains how Claude Code's auto memory stores notes as markdown files behind a MEMORY.md index, why that store fills up with duplicates, contradictions and stale dates, and how the /dream pass cleans it. He then rebuilds /dream as a custom skill from a publicly shared prompt. The explanation of the memory layout is clear and useful. His verdict that cleanup has no real downside is too relaxed: an automatic pass that merges and deletes memories needs version control and a human look at the changes.

The argument

The video argues that Claude Code's auto memory is helpful but degrades over time, and that a periodic consolidation pass fixes it. Auto memory writes small markdown files on its own when it hears something worth keeping, and lists them in a MEMORY.md index that loads at the start of every session. Left alone, the folder collects near duplicates, conflicting instructions, outdated facts and relative dates like "next Friday" that lose meaning, while the always-loaded index grows. Dream, which Anthropic was rolling out to a small group as of the March 2026 upload, reads the memory, checks it against recent session transcripts, merges and corrects files, and prunes the index. Since the prompt has been shared online, Chase turns it into a /dream skill with project, user and all scopes, and runs it on his own vault.

Key ideas, with timestamps

  1. 0:49

    Auto memory, not CLAUDE.md

    He separates auto memory from CLAUDE.md. Auto memory is what Claude Code writes by itself across sessions. His example: mention that you always go to the gym on Tuesdays, and a markdown file appears recording it without you asking.

  2. 1:24

    Where the files live

    The memory sits in the .claude folder under projects, then the project name, then a memory folder, not in your repo. Inside are individual markdown files plus MEMORY.md, a master index that links to each one, much like a skills list with descriptions.

  3. 2:23

    Index first, details on demand

    MEMORY.md loads at the start of each session. When a topic comes up, such as testing, Claude sees the index entry and opens the detail file, in his case a note saying he does not want mock tests and prefers real integrations.

  4. 3:03

    How it degrades

    Say "always use React" one day and "never use React" the next, and both files stay, both listed in the index. The folder keeps growing, and because the index is always loaded it adds context bloat. Relative dates (4:08) are a subtler problem: "next Friday" is stored as written and becomes wrong a week later.

  5. 4:30

    Dream's four steps

    Read the existing memory. Read recent session transcripts, the JSONL files in the same projects folder that log every message and tool call, roughly the last five sessions. Merge and fix files, including turning relative dates into absolute ones. Prune the index, which he says is capped at 200 lines and should be kept as short as possible.

  6. 6:43

    Rebuilding it as a skill

    The dream prompt was shared publicly by people with access; he credits a version from PyBites. You paste it into Claude Code and ask it to create a skill called dream. His addition is scope flags: plain /dream runs on the project memory, /dream user on the user-level memory that applies to every project, and /dream all on both.

  7. 7:51

    The demo run

    On his vault it flags seven issues: near duplicates, a contradiction, two stale entries, a relative date, a code convention that does not belong in memory, and over-verbose entries. Claude asks permission to edit its own .claude folder even with permissions bypassed. Result: a couple of files merged, four updated, three pruned, five unchanged.

  8. 8:29

    His verdict

    He calls it a marginal improvement rather than a change in how you work, but says there is no real downside to tighter, current memory. The video closes with plugs for his community and paid course.

What it gets right

The memory layout is explained well. Many users do not know auto memory exists, let alone where it lives. Knowing the path, that MEMORY.md is always loaded and that detail files are opened on demand is enough to inspect and edit your own memory by hand, which is the most useful takeaway here.

The failure modes are real and specific. Duplicates, contradictions, stale facts and relative dates are exactly how any append-only memory store rots. Relative dates in particular are easy to miss and cheap to fix: store "2026-03-27", not "next Friday".

The convention flag is a good sign. One of the seven issues was a coding convention sitting in memory. Conventions belong in CLAUDE.md, which you write and review on purpose; memory should hold things Claude learned about you. A cleanup pass that enforces that split is useful.

Where it is too optimistic

"No real downside" is not true. Consolidation is an LLM editing its own memory. When it resolves a contradiction it picks a winner, usually the more recent statement, and that can be wrong: "never use React" may have been about one project. When it prunes stale data it deletes, and you will not notice what is missing until Claude forgets something. The fix is simple. Put the memory folder under git (or copy it) before each run, and read the diff afterwards. Chase shows the summary counts, not the actual changes.

A third-party prompt is not the official feature. The skill is built from a prompt shared online. It may match what Anthropic ships, or an older or partial version of it. Treat it as a useful cleanup prompt, not a faithful copy.

Reading transcripts costs tokens. Step two reads several full session logs, which can be large on long coding sessions. Running it after every session is wasteful; weekly, or when the index gets long, is plenty.

Memory stays per tool. Everything here lives in Claude Code's own folder. The preference about mock tests is invisible to ChatGPT, Cursor or a Claude chat in the browser. As of the March 2026 upload, auto memory is a single-tool feature.

How this applies to memory across AI tools

Dream is a small version of a general rule: memory that is only ever appended to gets worse, so any memory system needs consolidation, conflict resolution and dates. The same applies to what you save from outside, such as videos, articles and papers, where the problem is less contradiction and more volume.

It also shows the limit of tool-local memory. If you work across several assistants, it helps to keep two layers apart: short, reviewed instructions per tool (CLAUDE.md and its equivalents), and a shared store of source material that any assistant can search. BrainTube covers the second layer, saving YouTube videos with their transcripts, articles, PDFs and podcasts into one compiled knowledge base exposed as an MCP server, so Claude, ChatGPT, Cursor and other MCP clients can all search it.

For more on the options, see Claude Code memory tools compared and persistent memory across tools.

Who should watch it

Watch it if you use Claude Code daily and have never looked at its auto memory folder, or if your sessions start with odd or outdated assumptions. The layout and failure modes from 0:49 to 4:30 are worth the full watch, and the skill setup and scope flags from 6:43 to 7:48 are what you need to copy. Skip 6:22 to 6:43 and the end card, which are course plugs. If you already manage memory by hand or keep everything in CLAUDE.md, there is little new here beyond the relative-date tip.

Frequently asked

Where does Claude Code store its auto memory?
According to this video, auto memory is stored in the .claude folder in your home directory, under projects, then the project name, then a memory folder. It holds individual markdown files plus a MEMORY.md index that loads at the start of each session. There is also a user-level memory that applies to all projects.
What does Claude Code's dream feature do?
Dream is a memory consolidation pass. It reads the existing memory files and recent session transcripts, merges duplicates, resolves contradictions, removes stale entries, converts relative dates to absolute ones and shortens the MEMORY.md index. As of March 2026 it was available to a small group of users.
Is it safe to let Claude clean up its own memory?
Mostly, but it can delete or overwrite correct memories when it resolves a conflict or prunes old entries. Back up the memory folder or put it under git before each run, then review the changes. Running it weekly rather than after every session also saves tokens.

Keep videos like this one searchable

BrainTube saves YouTube videos with their transcripts into one compiled knowledge base that Claude, ChatGPT, Cursor and other MCP clients can search.

Start free

Go deeper

More video notes

Are you the creator and want this page changed or removed? Contact us and we will act within 48 hours.