Search past sessions
search_content finds past sessions by text, on your own machine. Your agent asks for a session the way you would, and gets the hits back with their titles, agents and dates.
search_contentthrough crbon-mcpYour coding agents start every session from zero. AI Memory keeps your past work searchable, on your own computer.
Rowan Ellis
Platform engineer, Tallyfold
110 of 110 lit
Sessions remembered
110
120 days, every agent
Coding agents
5
Claude Code, Codex, Gemini CLI
Projects
4
payments-api, ledger-web
Oldest session
120days
8 Jul to 4 Nov 2026
Session graph
Paid: agents search every session. Drag a session, hover for its card, click the graph to zoom.
A made-up engineer at a made-up company. No customer data.
Works with the agents and assistants you already use
Claude Code, Codex, Gemini CLI, Cursor and OpenCode get the memory tools through crbon-mcp. ChatGPT, Claude and Gemini chats are kept in the browser.
You choose which agents get the tools, duringcrbon installor with crbon memory agents.
Rowan asks Claude Code how the duplicate refunds were stopped. search_content returns three sessions from payments-api, two of them from other agents. get_messages reads around the second, and the answer arrives with the date and the change.
Plays once as it scrolls into view. Replay runs it again; reduced motion shows the finished transcript.
Rowan’s sessions
Scroll to play
Claude Code11 Sep
1. Duplicate refunds after webhook retries
Codex12 Sep
2. Add idempotency keys to the refund webhook
Cursor12 Sep
3. Unique index on provider event id
Three tools through crbon-mcp and a few commands for the rest. Local by default; the two paid switches are separate choices, each off until you turn it on.
search_content finds past sessions by text, on your own machine. Your agent asks for a session the way you would, and gets the hits back with their titles, agents and dates.
search_contentthrough crbon-mcpget_messages reads the messages around a result, so the agent sees the decision and what led to it. get_memory_status checks memory is ready.
get_messagesthrough crbon-mcpWrite one session out as Markdown, ready to paste into another agent, or as JSON or the raw transcript.
crbon memory export <session-id>crbon install asks whether to turn memory on and which agents get the tools. Change the list any time; crbon status shows it.
crbon memory agentsOnly the folders you list are archived, and the memory tools show only their sessions.
crbon config set sync-folders ~/work/acmeSummaries and search by meaning, built by an AI provider through Vercel’s AI Gateway. Crbon does not store the text, and semantic off stops it.
crbon memory semantic onUpload the prompts and replies from chosen folders. Search them, read transcripts and see your graph. On Free and the paid plan only you see them; workspace admins do not.
crbon memory sync onBrowser Chat Memory
ChatGPT, Claude and Gemini chats stay on your device, in the browser. Open the drawer in Chrome or Edge and pick up any captured conversation.
Back up a chat you choose, add labels to backed-up chats and search local and backed-up chats together. A backed-up chat stays searchable after the browser drops its local copy.
Chat Memory
On this deviceExplain idempotent webhook handling
12 SepA retry can deliver the same refund event twice, so the handler needs a key it can check before it acts.
Draft the incident note for the refund outage
14 SepThe handler now stores the provider event id under a unique index and skips any repeat.
Backed uppayments
Compare Postgres unique index options
12 SepA partial unique index on the refund provider id keeps the table small and the check fast.
Summarise the refund policy for support
28 OctThe window for a customer refund is 30 days from capture, with partial refunds allowed once.
Attach adds the chat as a file to the other assistant’s box. You review and send.
Free and usage-based
| Memory feature | Free | Usage-based |
|---|---|---|
| Coding agent history | Newest 200 sessions | Every session |
| Search sessions by text | Included | Included |
| Read around a hit, export, agent choice, folder list | Included | Included |
| Search by meaning and summaries | Not included | Optional, with crbon memory semantic on |
| Dashboard sync and session graph | Not included | Optional, for sessions in chosen folders |
| Browser Chat Memory | Local capture, search and export | Plus per-chat backup, labels and combined search |
$1.99 once, then $0.10 per million billable tokens. There is no monthly fee. Full agent history comes with it; search by meaning, sync and chat backup are separate switches, each off until you turn it on.
Agents search your newest 200 sessions on your own machine. Browser chats stay on the device with search and export.
Everything the plans includePrivacy, plainly
On your computer. Coding sessions are indexed under ~/.crbon/memory and stay there by default. Browser chats stay on the device, in the browser. A folder list limits which folders are archived and which sessions the tools can show.
No session text leaves your machine until you turn a paid switch on. The CLI’s usage meter sends usage metadata, never your prompts. Semantic search sends session text to Crbon’s endpoint, then to an AI provider through Vercel’s AI Gateway to build the index; Crbon does not store that text. Sync uploads the prompts and replies from chosen folders, with home paths made ~ and known secrets scrubbed first.
Crbon stores synced sessions and can read them; they are not end-to-end encrypted. In the dashboard only you see them. On Free and the paid plan, workspace admins do not. crbon memory sync off stops new uploads, and sync delete --yes removes the cloud copy.
These promises cover Free and the paid plan. Enterprise Corporate Memory, below, is in development and changes none of them today.
Enterprise Corporate Memory is in development and not available yet.
Today memory is personal. On Free and the paid plan only its owner sees it, and workspace admins do not. Enterprise would add the views and the sharing an organisation needs, under a policy the organisation sets.
An admin can open any member’s AI memory, one person at a time.
Admins can share a member’s memory with colleagues when the work needs it, for example handing work over when someone changes role or leaves.
Supervisors see the memory of their direct reports and their group, and can share it.
Members can share their own memories with any colleague.
Set by the AI admin, including whether members can opt out.
Groups such as HR and Finance, kept apart from everyone else.
Filters that keep personal and non-work conversations out of company memory.
One graph across the organisation that shows how work connects within and between teams, and finds overlapping and duplicated effort.
In development, not available yet
Sample organisation
Tallyfold, made up
Larger nodes are people; small ones are their sessions. Dashed magenta threads join work across Payments, Platform and Data. Nothing crosses the Finance ring.
Get started
Install, answer yes to memory, choose your agents. Then ask about the work you were doing.
Read the CLI docsInstall version 0.2.4 of @crbonfree/cli from npm.
npm install -g @crbonfree/cli
Run the installer, answer yes when it asks about memory, then choose the coding agents that should receive the tools.
crbon install
Confirm memory is ready, then ask your agent to find an earlier session and read the messages around the result.
crbon status
Find my last session about the flaky checkout test and read around the decision.
Useful later
crbon memory agentscrbon memory offcrbon config set sync-folders ~/work/acmecrbon memory export <session-id>Questions
Local paths, plan limits, cloud choices and deletion, without the fine-print routine.
Read the CLI docsCoding memory works with Claude Code, Codex, Gemini CLI, Cursor and OpenCode. You choose which agents receive the tools during crbon install or with crbon memory agents. The Chrome and Edge extension captures chats from ChatGPT, Claude and Gemini.
They are indexed on your own machine under ~/.crbon/memory and stay there by default. With crbon config set sync-folders ~/work/acme, only the folders you list are archived and only their sessions appear through the memory tools.
Free lets agents search your newest 200 sessions and includes local browser chat memory. The paid plan covers every agent session and adds opt-in semantic search, opt-in sync and browser chat backup features. It is usage-based: $1.99 once, then $0.10 per million billable tokens.
On the paid plan, crbon memory semantic on enables search by meaning and summaries. Session text goes to Crbon’s endpoint, which forwards it to an AI provider through Vercel’s AI Gateway to build the search index. Crbon does not store that text. Run crbon memory semantic off to stop it.
crbon memory sync on uploads prompts and agent replies from sessions in your chosen folders to your CrbonFree dashboard. Home paths become ~ and known secrets are scrubbed before upload. Crbon stores synced sessions and can read them; they are not end-to-end encrypted. Only you see them in the dashboard, not workspace admins. Run crbon memory sync off to stop new uploads.
Filter by assistant, search titles and messages with highlighted matches, or search within one conversation. You can copy, rename, open the original chat or export it as .md, .txt or .json. Attaching a chat to another assistant adds a file to its chat box; you review and send it yourself.
crbon memory sync delete --yes deletes the cloud copy. Add --session <id> to delete one synced session. Turning sync off stops new uploads without deleting what is already there.
Memory is free on your computer. Cloud features are paid and opt-in.