Why Claude Keeps Forgetting You (And How To Fix It)

Overview

The video explains why Claude (and similar AI chatbots) “forgets” and introduces a 7-level context-management system designed to help an AI reliably remember you and your work.

The core idea is that context (what the model knows before you prompt it) strongly determines answer quality—while chat-based context is limited by the model’s context window.

---

Level 1 — Chat history (ask it to “catch up”)

• Claude/ChatGPT can read previous conversations if you ask (for example, start a new chat and instruct it to continue where you left off).
• As chats get longer, they consume more of the context window, degrading output quality.
• Starting fresh avoids an ever-growing thread, but you still don’t control what gets selected from history.

---

Level 2 — Upload documents (point-in-time references)

• Upload Word documents, PowerPoints, and PDFs to provide exact context.
• Key issues:
  1. Point-in-time references: the AI can read uploaded files, but it can’t “save” or update them automatically—so you must re-upload for new chats.
  2. Cost/performance: these are binary files, so the system typically must do text extraction first, which is slower and uses more tokens.

---

Level 3 — “Projects” (persistent context inside an app)

• Create a project in Claude/ChatGPT, load documents, and then do all work within that project to reuse the same context.
• Benefit: context is more organized and can be shared across a team.
• Limitation: the context lives inside a single vendor’s product, making it hard to switch tools/providers without losing access to the stored setup.

---

Level 4 — Context files (local markdown “second-by-design” memory)

• Use markdown files (text with formatting) as context sources, such as:
  • about.md
  • voice.md
  • memory.md
• Advantages vs PDFs/Word:
  • No binary extraction step
  • Cheaper/faster
• Limitation: if placed only inside a chat, they still require manual upkeep.

Key unlock: AI coding/work tools that read/write files locally

The video emphasizes using tools (examples mentioned: Claude Code, CodeX, Co-work, ChatGPT Work) that can read, search, and rewrite context files on your machine—turning them into living documents rather than static uploads.

Example starter setup

• about.md: who you are / what you do / who you sell to
• voice.md: writing style guidelines
• memory.md: tracked preferences over time

The claim: once these files are stored locally, you can switch AI tools later (as long as the tool can access the files).

---

Level 5 — “Second brain” (a context library folder)

A second brain is essentially a folder/library of context files (often called LLM wikis or context layers).

• Simplest form: markdown files in folders with minimal metadata and links.
• Guidance: keep it simple—less contradictory context beats thousands of messy files.

Maintenance

• Linting is described as a scheduled process/agent that organizes and consolidates the library weekly to prevent context drift.
• Mentions a live training/event: “Founder OS Build Shop” on August 26.

---

Level 6 — Context farming (autonomous ingestion/updating)

The second brain updates itself via scheduled/triggered agents that fetch new sources and ingest them.

• Examples of automation:
  • Agents monitor creators for new posts
  • Agents ingest new sales-call transcripts automatically
• Constraints:
  • Users choose the curated sources (“you decide what to monitor”)
  • Agents only check for updates; they don’t freestyle content
• Emphasizes compounding effect when combined with linting.

---

Level 7 — Company brain (scalable, permissioned knowledge base)

For organizations, a “company brain” adds to the personal second brain concept:

• Structured storage
• Permissions

Architecture described

• Use a database (typically Postgres) with one table.
• Each row represents one knowledge item (e.g., a Slack thread, code snippet, or doc section).
• Each row stores:
  1. the content
  2. an embedding (content converted to numbers for semantic search)
  3. metadata (source + permissions)

Retrieval mechanism

• The AI connects via MCP or a connector to search the database for relevant context.

Scaling note

• Sharing second brains works up to ~15 people, but beyond that, retrieval/structure breaks down.

Permissions emphasis

• Company brain ensures only authorized employees can retrieve sensitive knowledge.

---

Main speakers / sources

• Single primary creator/narrator (the video author), speaking throughout.
• Mentions/tools as sources/examples:
  • Claude
  • ChatGPT
  • Work tools like Claude Code / CodeX / Co-work / ChatGPT Work
• Mentions “Founder OS Build Shop” (event/training) as the creator’s related product/source.