Skip to contents

graft

Project memory for R apps and agents

Give your next conversation
the knowledge your team has reviewed.

Save a finding or an agreed definition with the evidence behind it. An agent or workflow can look it up later, check its exact sources, and see how it changed.

Build a project memory assistant Get the runnable Shiny example

Remember an agreed definition

You’re building an analytics assistant with ellmer and shinychat. The team agrees that an active customer has a paid account and used the product in the last 30 days. Next week, someone opens a fresh conversation and asks how to count active customers.

With Graft, your app can retrieve the reviewed definition and the exact source behind it. If the team changes the window to 60 days, you can review that correction, use it in later conversations, and still inspect the earlier version.

Review once Reuse later Correct with history
Save the agreed definition and its handbook excerpt. Let a fresh chat retrieve the current reviewed note through graft_tool(). Review the updated definition while preserving the original evidence.

Try it in a Shiny app

The included project notebook puts review controls beside a shinychat conversation. The default scripted preview uses real Graft files without model calls.

pak::pak(c("JamesHWade/graft", "ellmer", "shinychat", "shiny", "bslib"))
Sys.setenv(GRAFT_DEMO_STORE = file.path(getwd(), "project-memory"))
shiny::runApp(system.file("examples", "project-memory", package = "graft"))

Review and save the proposed definition, then ask “How do we count an active customer?” Start a new conversation and ask again. Stop and restart the app to reopen the same notebook. The getting-started guide walks through the storage calls, connects graft_tool() to ellmer, and explains how to enable the live agent.

Where you could use Graft

Project knowledge for an agent

Save a reviewed definition, decision, or finding for another conversation. Your app chooses what to keep and who can use it.

Give a chat agent project memory

A report with its evidence

Keep a conclusion, table, or figure with its exact inputs so you can check what went into the result.

Keep artifacts and review history

Shared concepts across workflows

Save definitions and relationships with a specific data-dict release so agents and applications can recover the same vocabulary.

Publish shared vocabulary

Save a result with its evidence

Start with a fresh directory and ordinary text:

library(graft)
store <- graft_store("analysis-memory", create = TRUE)
source <- graft_save(store, "Handbook: use a 30-day activity window.", "handbook")
note <- graft_save(
  store, "An active customer used the product in the past 30 days.",
  id = "active-customer", dependencies = source
)
graft_read(store, note)@data

Reopen it later with graft_store("analysis-memory"). Saving a correction under the same ID creates a new revision; the original reference still reads the original content. Use graft_accept() to record a review and graft_recall() to retrieve the current accepted result, as shown in the getting-started guide.

Add Graft to an existing application

Keep ellmer for the model and tool calls, shinychat for the conversation, and your application’s review and access controls. Add Graft where a useful result should outlive the current chat or R process.

Start with one kind of note and a read-only graft_tool() fixed to its stream and purpose. A larger app can use its own catalog or search index to choose relevant notes, then ask Graft to verify the saved selections and evidence.

Storage and operational guides

A local Graft store uses trusted files, and several R processes can write to it on a file system that honours advisory locks (many network file systems do not). For PostgreSQL, your application owns the transaction in which Graft reads and writes. This pre-production example is for one trusted local user. Before deploying to several users, your app needs to check access to each user or project and decide how long to keep the data.

Permanent Forget and recovery from retired backups remain active work. Saving or reviewing material records your application’s decision; it does not establish that a claim is true.