# SOURCE: Claude Memory Boundaries — Public Demo Research — October 8, 2026 ## Multiplist Source Delivery Document | ID: 923aa15c --- **Type:** import | **Chars:** 3,360 | **Seeds:** 28 | **Updated:** 2026-10-08 **Status:** complete | **Assets:** 0 --- ## Abstract This document consists of research notes detailing Claude's memory boundaries and context management mechanisms, prepared for a Multiplist public demonstration. It clarifies the distinctions between project, chat, and Claude Code memory, emphasizing that various features often operate with separate memory scopes, especially for local Cowork sessions. The notes also address how file uploads are handled, the ongoing merger of Claude's chat and Cowork experiences, and propose a strategy for explicit portable context handoff. The overall purpose is to provide factual, documented information to guide users and manage expectations regarding memory continuity across different Claude functionalities, particularly in the context of an evolving product experience. --- ## Decisions (3) **[D1] Notes based on public documentation for Multiplist demo** (confidence: high) These are original research notes based on public documentation, prepared for a Multiplist demonstration. ↳ Lines 2–2 **[D2] No private customer or founder context** (confidence: high) They contain no private customer or founder context. ↳ Lines 2–2 **[D3] Demo does not stage production screen or claim upload** (confidence: high) It does not stage a production generation screen or claim a tested upload into a different AI account. ↳ Lines 28–28 --- ## Frameworks (1) **[F1] Claude Code context mechanism** (confidence: high) Claude Code carries context through instruction files such as CLAUDE.md and through auto memory. ↳ Lines 17–17 --- ## Golden Passages (15) > "Each project maintains separate memory and a project summary." > ↳ Lines 2–2 > "Chat memory can carry into cloud Cowork tasks, while computer-local Cowork sessions do not use that shared memory." > ↳ Lines 2–5 > "Check the execution environment and memory settings before expecting continuity." > ↳ Lines 5–5 > "The reader should distinguish project context from account-wide continuity." > ↳ Lines 6–8 > "Anthropic is gradually merging chat and Cowork into a single conversation experience, initially on Pro and Max. Accounts can be at different rollout stages." > ↳ Lines 9–9 > "A universal claim that chat and Cowork never share memory is outdated." > ↳ Lines 9–10 > "Its auto memory is machine-local and shared across worktrees of the same repository. This does not establish automatic synchronization with Claude chat memory." > ↳ Lines 10–13 > "Uploading a file does not promise that every embedded record will enter persistent memory." > ↳ Lines 13–14 > "For a portable handoff, preserve the research question, concise answer, evidence, review date, and unresolved questions in a file." > ↳ Lines 14–17 > "A dated evidence snapshot remains useful when provider interfaces change." > ↳ Lines 17–17 > "It does not automatically synchronize accounts or certify the truth of source claims." > ↳ Lines 17–18 > "The shareable demonstration uses public research notes, canonical Neuron extraction, a client-authored synthesis, and the existing v3 HTML capsule renderer." > ↳ Lines 18–21 > "Creating Archivist briefs is a Plus feature; reading a shared static HTML file requires no Multiplist subscription." > ↳ Lines 24–25 > "The download carries the brief and complete selected evidence records, not the full upstream documents." > ↳ Lines 25–25 > "They are evidence about documented behavior, not a live test of a Claude account" > ↳ Lines 2–2 --- ## Definitions (3) **Distinction between project and non-project chats:** Anthropic's memory guide distinguishes project memory from non-project chats. ↳ Lines 5–5 **Definition of a project:** A project combines files, instructions, context, and memory. ↳ Lines 9–9 **HTML as supported document upload format:** Anthropic lists HTML as a supported document upload format. ↳ Lines 21–21 --- ## Actions (2) - **A reader should be able to inspect the supporting text before reusing the answer.** ↳ Lines 25–25 - **Review included text before sharing.** ↳ Lines 28–28 --- ## Offers (1) - A recipient can attach a brief capsule as reference material. ↳ Lines 21–21 --- ## Emergence (3) **[E1] The necessity of explicit memory scope awareness** (confidence: high) The document repeatedly emphasizes the need for users to actively understand and distinguish between different memory scopes within Claude (chat, project, Cowork, Claude Code). This consistent clarification acts as a guardrail against assumptions of automatic or universal memory continuity, underscoring a tension between integrated user experience and discrete technical boundaries. ↳ Lines 2–13 **[E2] Navigating a product in transition** (confidence: high) The content highlights that Claude is an evolving product, with features like chat and Cowork actively merging, but acknowledging varying rollout stages and 'outdated' universal claims. This creates an ongoing challenge for users to understand the current state of memory behavior versus anticipated future states, requiring constant vigilance of documented changes. ↳ Lines 9–10 **[E3] Proactive context management vs. automatic synchronization** (confidence: high) There is a strong pull towards explicit, user-driven context management, particularly through the described 'portable handoff' strategy. This contrasts with an implied expectation of seamless, automatic synchronization across various Claude features, which the notes consistently state does not occur by default for many memory types. ↳ Lines 13–18 --- ## Source Content # Claude memory boundaries — public demo research notes Reviewed: 2026-10-08. These are original research notes based on public documentation, prepared for a Multiplist demonstration. They contain no private customer or founder context. They are evidence about documented behavior, not a live test of a Claude account. ## Chat and project memory Anthropic's memory guide distinguishes project memory from non-project chats. Each project maintains separate memory and a project summary. Chat memory can carry into cloud Cowork tasks, while computer-local Cowork sessions do not use that shared memory. Check the execution environment and memory settings before expecting continuity. Reference: https://support.claude.com/en/articles/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context ## Project context A project combines files, instructions, context, and memory. A project tied to an existing local folder remains on that computer. The older Cowork project guide describes memory scoped to the project. The reader should distinguish project context from account-wide continuity. Reference: https://support.claude.com/en/articles/14116274-organize-your-tasks-with-projects-in-claude-cowork ## A changing product experience Anthropic is gradually merging chat and Cowork into a single conversation experience, initially on Pro and Max. Accounts can be at different rollout stages. The current guide says memory from computer-local Cowork tasks stays with those tasks. A universal claim that chat and Cowork never share memory is outdated. Reference: https://support.claude.com/en/articles/16761823-claude-cowork-and-chat-are-one-claude ## Claude Code is a separate context mechanism Claude Code carries context through instruction files such as CLAUDE.md and through auto memory. Its auto memory is machine-local and shared across worktrees of the same repository. This does not establish automatic synchronization with Claude chat memory. Reference: https://code.claude.com/docs/en/memory ## Reusing a file in Claude Anthropic lists HTML as a supported document upload format. A recipient can attach a brief capsule as reference material. Uploading a file does not promise that every embedded record will enter persistent memory. Reference: https://support.claude.com/en/articles/8241126-upload-files-to-claude ## Portable context as an explicit handoff For a portable handoff, preserve the research question, concise answer, evidence, review date, and unresolved questions in a file. A reader should be able to inspect the supporting text before reusing the answer. A dated evidence snapshot remains useful when provider interfaces change. It does not automatically synchronize accounts or certify the truth of source claims. ## Multiplist demonstration boundaries The shareable demonstration uses public research notes, canonical Neuron extraction, a client-authored synthesis, and the existing v3 HTML capsule renderer. The local video shows the actual exported file. It does not stage a production generation screen or claim a tested upload into a different AI account. Creating Archivist briefs is a Plus feature; reading a shared static HTML file requires no Multiplist subscription. The download carries the brief and complete selected evidence records, not the full upstream documents. Review included text before sharing. --- *Pulled via Multiplist MCP | 2026-10-15T14:05:37.777Z* *"Recursion compounds. Summarization collapses. Provenance is integrity."*