Your AI agent remembers what it is told, says where it learned it, and keeps every earlier version. When a memory goes wrong, you can see what changed and undo it.
A team tells its AI agent a few things. Each one is kept with where it came from, so the agent can always say how it knows.
00:30
Things change. Nothing is lost.
Later the release day moves, support hours get longer and the stand-up is cancelled. The agent updates its memory, and every earlier version stays in the history.
02:10
See exactly what changed.
Pick two moments and Bilinc lists every fact that was added, changed or removed in between, and where each one came from. Here: 21:00 against 02:10.
Like a long exposure from 21:00 to 02:10, every star leaves a trail. A fact's trail turns amber where it changed, starts in cyan where it was added and stops in red where it was removed.
Facts that changed between 21:00 and 02:10
Fact
21:00
02:10
The page is back at 02:10. Nothing was lost: scroll on to return.
03:40
Undo a wrong memory.
A forwarded email draft says the launch moved to 20 October. It didn't. Ask the agent now and it would announce the wrong date. Preview going back to 02:10, then undo it.
The wrong memory
The next chat
Someone asks: When is the launch?
1
Source 1, from recall
Your agent wants to roll memory back to 02:10
Hold until it fills, then let go. Letting go early cancels.
Cancelled. Nothing changed, and the wrong date stays.
06:20
Add it to your AI in a minute.
Connect Bilinc to Claude, Cursor, Codex or any MCP client at mcp.bilinc.space/mcp, signed in with OAuth.
pip install bilincInstalls the Python package and the bilinc command.
bilinc loginSigns you in through your browser and keeps the key on your machine.
bilinc quicktestWrites one fact and reads it back, so you know it works.
Your agent gets ten tools:
remember
recall
revise
forget
create_snapshot
list_snapshots
diff
preview_rollback
rollback
status
What Bilinc is.
Memory for AI agents that keeps its history.
Bilinc is a hosted memory your AI agent reads and writes while it works. When an agent learns something worth keeping, a decision, a preference, the outcome of a task, it saves it to Bilinc under a key. Later, in another chat, another session or another app, it recalls it instead of asking you again.
What makes Bilinc different is what it keeps around each memory. Every write records where it came from: the app or run that wrote it and when. A change never silently replaces what was there; the earlier value stays in the memory’s history. You can save a checkpoint of the whole workspace, compare any two moments, and roll the workspace back to a checkpoint when a run writes something wrong.
Bilinc Cloud is the hosted service. Agents reach it over the Model Context Protocol (MCP), from Python through the SDK, or from the terminal with the bilinc command. The bilinc package is source-available under BUSL-1.1 and talks to Bilinc Cloud; it does not ship a local storage engine.
Who it is for.
Anyone whose agent should remember, and who needs to know how it knows.
Developers with coding agents
Claude Code, Codex, Cursor and other coding agents recall the decisions a project already made before they edit, and keep what each session settled: what changed, how it was verified, what is still open.
People who use Claude or ChatGPT
Add Bilinc once and what you tell one app is there in the next chat, and in the other apps you connect to the same workspace. Each memory records which app wrote it.
Teams
A shared workspace gives a team’s agents one memory. Anyone can see where a fact came from and how it changed, instead of finding it overwritten.
Operators of agent systems
When a run goes wrong, you can answer which run wrote a piece of state, what it replaced, and how to roll it back, without building that record yourself.
Bilinc is not a vector database and not an agent framework. It is the memory your agent calls, and it works alongside the retrieval and orchestration you already have. Why vector memory isn’t enough explains the difference.
How it works.
The same night as above, in plain terms.
Versioned writes
An agent connected through the remote server saves a new memory with remember, which only creates new keys, so a new session cannot quietly replace what an earlier one stored. From the SDK and the local adapter, commit_mem writes a memory with its provenance and returns its version. To change a memory on purpose, the agent calls revise, which can require the version it last read, so a change made elsewhere is never overwritten unseen. Every earlier value stays in the memory’s history, newest first.
Recall
recall searches the workspace and returns each memory with its source, when it was last written or confirmed, and its version. Once a memory is past its review window it is marked as possibly outdated, and still returned: old is not the same as wrong. Confirming a memory that still holds refreshes its date and keeps the earlier state in history.
Diff
diff compares a checkpoint with the current state or with another checkpoint, and lists every memory that was added, changed or removed in between. That is the 02:10 scene above: one call shows what a night of work did to memory.
Snapshots and rollback
Before risky work, an agent or a person saves a checkpoint of the whole workspace. If a run goes wrong, preview_rollback shows what going back would change, and rollback restores the checkpoint with the token the preview issued. Rollback works on checkpoints, not on single keys, and in Claude it always asks you first.
Audit trail
Each write keeps the record you need to review it later: which app, agent or operator wrote it, where the value came from, and what it replaced. Removing a memory with forget requires a reason, and that reason is recorded. Signed-in members can browse a workspace’s memory read-only in the dashboard and export it as JSON.
Your control
When you connect an app, you choose the workspace and whether the connection is read-only or read and write. A read-only connection only sees the read tools. You can disconnect any app at any time under Connected apps in the dashboard.
One workspace, every client. Claude and ChatGPT connect to https://mcp.bilinc.space/mcp and sign in with OAuth; coding agents can also run the local adapter from the bilinc package.
Any client that can start a stdio MCP server or connect to a remote one can use Bilinc. From your own code, the Python SDK and the bilinc command use the same hosted memory. See how to connect over MCP.
Install it.
Free, no card. Pick the path that matches how you work.
In Claude, ChatGPT, Claude Code or Codex
Add the server. In Claude, open Customize → Connectors → Add custom connector. In ChatGPT, open Plugins, then Add → Add custom MCP server. The address is the same everywhere:Server URL
Sign in. The app opens bilinc.space (in Claude Code, run /mcp, pick bilinc and choose Authenticate). Sign in or create an account, pick a workspace, and choose read-only or read and write.
Ask. For example: “Check Bilinc: what did we decide for the billing database?”
For developers
Install the package and sign in. One package gives you the Python SDK, the bilinc command and the MCP adapter; bilinc login opens your browser and keeps the key on your machine.Terminal
pip install -U bilinc
bilinc login
Check that it works.bilinc quicktest writes one fact and reads it back.Terminal
bilinc quicktest
Connect your agent.bilinc mcp install prints the MCP server entry with your Python path, ready for Cursor, VS Code, Gemini CLI, Devin Desktop or any other client.Terminal
No. A vector database answers “what is similar to this?”. Bilinc keeps the record around each memory: where it came from, what it replaced, and how to undo it. The two can work side by side.
Which apps and agents can use Bilinc?
Claude on the web, desktop and mobile, Claude Code, ChatGPT, Codex, Cursor, VS Code, Gemini CLI, Devin Desktop, opencode, Hermes-Agent and any other MCP client, plus your own code through the Python SDK and the CLI.
Can I undo something my agent remembered?
Yes. Save a checkpoint before risky work; if a run writes something wrong, preview the rollback and restore the checkpoint. Rollback works on checkpoints, not on single keys, and every earlier value stays in history.
How long is memory kept?
Memory is kept on every plan with no time limit, and is not deleted when a plan lapses or an account returns to Free.
Can I read the source?
The bilinc package is source-available under BUSL-1.1. It is the client side: the Python SDK, the CLI and the MCP adapter that talk to Bilinc Cloud.
How do I start?
Create a free account, then add https://mcp.bilinc.space/mcp to your app or run pip install -U bilinc and bilinc login. The free tier needs no card.
