Use case

Continue an AI conversation across days, devices, and providers

restore_chat is the universal "pick up where you left off" tool: across providers, across devices, across team members. The portable manifest is what makes this work without the receiver's LLM needing to be the original conversational partner.

The three contexts where restore_chat shines

Across providers (Claude → ChatGPT, see /use-cases/cross-provider-handoff). Across devices (Macbook → iPad, where the chat history doesn't sync because Claude.ai doesn't sync). Across team members (Alice → Bob for a handoff). The portable manifest assembles the same three pieces — system prompt + memory snapshot + receipt digest — for all three cases. The receiver model picks up the work without a re-brief.

How to invoke it

In Claude Desktop / Cursor / Cline / Windsurf with GenZAgents installed: say "use the restore_chat tool to bring back my work on the Q3 forecast". The MCP server queries the receipt pool for matching past work, assembles the manifest, returns it as the assistant's reply. Continue the conversation from there. In ChatGPT without MCP: paste the portable-manifest output from /agents/[did]/portable into a new ChatGPT chat.

How fidelity holds up

Conservation of detail: the manifest is necessarily lossy (you can't fit a 3-hour conversation into a 4k-token system prompt). The summary LLM prioritises working assumptions and key decisions over verbatim turn-by-turn. In our design-partner measurements, ~85% of context that matters for continuation is preserved — sufficient for the receiver to act as if it had been there, except for fine semantic nuances that probably weren't the load-bearing parts of the conversation.

Why this doesn't exist natively in providers

Anthropic's memory feature is per-account; OpenAI's memory is per-account; Google's Workspace memory is per-Workspace-user. None of them work cross-provider because cross-provider portability hurts vendor lock-in. The neutral metalayer is the only place this can live — same logic as Plaid, Twilio, Stripe.

Combined with org_context_lookup — the proactive layer

org_context_lookup fires at conversation start automatically. restore_chat is the explicit "bring back X" command. Together: the system surfaces past work proactively when you start a fresh conversation, and gives you a precise tool to fetch a specific past conversation when you want it. Most users use both without thinking about which is which.

When restore_chat doesn't help

Brand-new conversations with no past receipt history. Conversations whose original context was a niche tool (proprietary internal API, custom system prompt with secrets that can't leave the original session). For the latter, the receipt is still preserved for audit; just the restore-as-continuation isn't the right fit.

Common questions

How do I find a past conversation to restore?

Three ways: search by project tag, search by free text (hybrid lexical + semantic), or browse the agent's receipt feed at /agents/[did]. Each receipt has a "restore this chat" button.

Can I edit the manifest before sending it to the receiver?

Yes — the manifest is JSON, editable in any text editor. Strip sensitive fields, add additional context, narrow the focus. The receipt of restoration is unchanged regardless.

Does restoring a chat create new receipts?

The continuation conversation creates new receipts as it runs. The restore action itself is logged as an audit event (not a new working receipt) so the audit chain is complete.

What's the max age of a receipt I can restore from?

No hard limit — receipts are stored indefinitely on paid plans. Free plan keeps 90 days. Practical upper bound: very old receipts may have stale context that the model can't usefully build on.

Related

Get the trust layer for your AI work

GenZAgents is the verified work-history layer above every AI provider your team uses. Sign cryptographic receipts, hand off conversations across Claude / ChatGPT / Cursor / Gemini, keep institutional AI knowledge when employees leave.

Last reviewed · 3 min read· Open spec· Changelog