Claude Cross-Session Memory: Chat, Code and Shared Vaults

John Rood··9 min read

Yes, Claude has cross-session memory. Claude chat can remember context and search earlier chats. Claude Code carries instructions and auto memory into new sessions. Neither is the same as reopening a complete conversation, and neither automatically gives another AI tool your conversation history. Choose the layer based on what you need to carry forward.

Checked November 9, 2026 against Anthropic's current documentation. Claude's newer memory experience and its legacy summary-based experience have different controls; check which one your account shows.

What persists across Claude sessions?

Memory layer What a new session can use Boundary to understand
Claude chat memory Saved topics about your work, preferences and ongoing context Account and project settings control what is used; this is not a complete transcript
Claude past-chat search Relevant earlier conversations retrieved when Claude searches Search and saved memory are distinct features with separate controls
Claude project memory The project's memory plus account memory when enabled Other projects' memory is not automatically included
Claude Code instructions Rules you write in CLAUDE.md or supported AGENTS.md files Instructions are not a record of every conversation
Claude Code auto memory Notes Claude writes about corrections, preferences and project context Per-repository notes, not a shared cross-tool conversation archive
MemoryRouter Relevant conversation memory from the connected vault Claude chat uses connector tools; Claude Code can use automatic hooks

Sources: Anthropic's Claude chat search and memory guide, Claude Code memory documentation and session behavior documentation. MemoryRouter behavior is documented separately for Claude connectors and Claude Code.

Claude chat memory: enable it before adding another layer

In Claude, open Settings > Your Claude, or Settings > Memory in the older layout. Check Generate memory from chats and Search and reference chats. Use Manage next to your memory to inspect, edit or delete what is saved.

Anthropic's current guide describes memory as individual topics saved while you chat, rather than only a summary generated afterward. You can ask Claude to remember a decision or correct an existing memory. The same memory is available to Claude chat and Cowork in the cloud. The guide also documents the legacy experience, where a memory summary updates periodically. Do not assume both accounts have identical timing or controls.

For a practical check, ask Claude to remember a harmless project preference. Start a new chat and ask what it remembers. Then inspect the saved topic or memory summary. If you need an exact detail from an earlier conversation, ask Claude to search that conversation and inspect its citation instead of treating a plausible answer as proof.

Incognito chats are not saved to chat history or memory. Turning memory off for a normal chat is different: that chat stays in history. Review Anthropic's controls and exclusions before using memory for sensitive work.

Claude cross-project memory: check the account-memory switch

In the newer experience, each Claude project has its own memory. While you work inside a project, Claude uses that project's memory alongside account memory from chats outside projects, but not memories from other projects. New project memories can also be referenced outside projects when memory is on.

If a project should stay separate, turn off Use account memory under Memory on the project's page. The legacy experience has separate project summaries and boundaries, so use the current project-memory guidance for your account rather than assuming that a project name alone establishes isolation.

This is a reason to configure native memory carefully, not a reason to disable it. Add another system when the requirement is different: the same conversation memory available across Claude Code, another repository, or another connected AI tool.

Claude Code cross-session memory: resume, instructions or recall?

Claude Code does have memory across sessions. Its native mechanisms are CLAUDE.md instructions and auto memory. Claude writes auto memory notes itself; the first 200 lines or 25 KB of MEMORY.md, whichever comes first, load at session start. Topic files can be read when needed. These notes are useful persistent context, not a promise that every decision and conversation detail was captured.

Choose the smallest tool that fits:

  • Continue the exact unfinished task: use claude --continue or claude --resume to reopen the earlier conversation. Anthropic documents that a genuinely new session starts with a fresh context window, while resume reopens saved history.
  • Apply the same rules every session: put stable instructions in CLAUDE.md. Let auto memory retain useful corrections and project learnings.
  • Retrieve conversation context across fresh sessions, repositories and tools: connect a shared memory vault and verify the details you need were actually captured.

For the native mechanisms, see how Claude Code remembers projects. For resume and the difference between saved history and a new session, see how Claude Code sessions work. Our Claude Code setup page covers the additional MemoryRouter layer.

One MemoryRouter vault across Claude surfaces and other tools

MemoryRouter gives connected tools persistent conversation memory outside their context windows. A decision captured in Claude Code can be retrieved in a Claude browser chat connected to the same vault. Other supported tools can use that vault too. This does not synchronize Claude's native saved topics or silently import old chat history.

The connection method changes the behavior:

Claude chat and Cowork: explicit connector recall and saving

Anthropic supports custom connectors using remote MCP. In Customize > Connectors, add a custom connector using:

https://mcp.memoryrouter.ai/mcp

Follow Claude's OAuth connection flow, sign into MemoryRouter and choose the intended vault. Managed workspaces may require an owner to add the connector first. Do not put a Memory Key or provider key in an OAuth client-secret field. Enable the connector for the conversation through + > Connectors. The connector installation guide covers authentication and verification.

Ask Claude to search MemoryRouter when earlier context matters. Ask it to save a concise decision, or approve the exact memory it proposes before it writes. These are model-directed tools: connecting them does not guarantee automatic recall or capture on every turn. Claude's native memory remains a separate system.

Claude Code: automatic recall and turn capture through hooks

From the project folder, install the MemoryRouter hooks:

npx -y memoryrouter-claude init

The installer asks for your Memory Key and checks the vault binding. The hooks retrieve relevant memory before prompts and capture the typed user prompt plus the final user-visible assistant response. They exclude tool payloads, reasoning and subagent chatter. Memory stays in the vault across sessions and context compaction, although recall retrieves relevant material rather than replaying the whole archive.

Run npx -y memoryrouter-claude doctor and follow the Claude Code verification guide. Installation starts capture going forward; it does not backfill earlier sessions. Installation scope controls where hooks load, not a separate memory partition per repository: the hooks use the whole configured vault.

Sharing company context without mixing everything

The unit to reason about is the vault, not a chat title or folder name. Only connect people and tools authorized to use that vault. Decide which information belongs in the shared collection before saving it. Do not copy a personal Memory Key into a team chat or repository.

Repository tags help identify captured work; they are not access-control boundaries. Use a separately authorized vault when context must be strictly isolated. A shared vault also does not turn every teammate's Claude conversation into company memory automatically: connector saves still require an explicit request or approval, and coding capture requires installed hooks.

Prove cross-session memory with a fresh conversation

Use synthetic information before relying on a new connection for real work:

  1. In a connected Claude chat, ask it to save: "The synthetic Harbor demo project's release phrase is ORCHID-7419." Approve the write and confirm store_memory succeeded.
  2. Open a genuinely new conversation with the same MemoryRouter connector enabled.
  3. Ask: "Search MemoryRouter for the Harbor demo project's release phrase." Do not repeat the answer or paste the earlier chat.
  4. Inspect the search_memories result for ORCHID-7419. A confident sentence without that tool result does not verify the connector.
  5. Repeat the search from another supported tool connected to the same vault to verify the cross-tool path.

If recall fails, check the vault binding, whether the save succeeded and whether the connector is enabled. For Claude Code, check doctor before assuming the model ignored a memory.

Frequently asked questions

Does Claude remember previous sessions?

Yes, when its memory features are enabled, Claude can carry saved context into new conversations and search previous chats. This is selective memory and retrieval, not every earlier message loaded into every new conversation. Check your saved memory and relevant search citations.

Does Claude Code have memory across sessions?

Yes. Claude Code loads persistent instructions and auto memory into new sessions. Resume an earlier session for its saved conversation history. Add MemoryRouter hooks when you want relevant captured conversation memory across fresh sessions, repositories and connected tools.

What is the best way to carry context between Claude Code sessions on one project?

Resume the old session if you want to continue that exact task. Use CLAUDE.md for standing rules and auto memory for retained learnings. Use a conversation-memory layer when the missing context spans several sessions or tools. These approaches complement each other.

Does Claude memory work across projects?

The newer experience combines a project's own memory with account memory when enabled, but does not automatically include other projects' memories. Check Use account memory for project separation. Legacy accounts have different summary-based boundaries; consult the linked Anthropic guide.

Can teams share one company memory across Claude, ChatGPT and Cursor?

MemoryRouter can provide a shared vault to authorized connections across supported tools. Connect each tool to the intended vault and verify a save-and-recall round trip. Capture and recall behavior depend on the integration; a connector is not an automatic transcript recorder or a replacement for permissions.

Is importing Claude memory the same as keeping tools connected?

No. Anthropic supports importing and exporting Claude memory, which can help move saved context between services. That is a transfer, not ongoing synchronization. MemoryRouter is a separate shared vault that connected tools can continue saving to and searching.

Create your MemoryRouter account, connect the Claude surface you use, and verify one saved decision in a fresh conversation.