grag documentation
Project memory that carries forward¶
Keep code, documents and decisions in a local graph your agent can use across sessions.
Set up your first project Explore the guides
What would you like to do?¶
-
Get connected
Install grag, connect your agent and verify a first read and write.
-
Navigate a codebase
Index selected sources and find definitions, callers and source citations.
-
Keep useful memory
Save decisions, correct earlier notes and resume unfinished work.
-
Find the right context
Retrieve relevant evidence within a budget and check its freshness.
-
Resolve a problem
Diagnose connection errors, missing libraries and database ownership.
-
Look up an interface
Find MCP tools, CLI flags, environment settings and write contracts.
A small local setup¶
Use one database per checkout, with multiple agent clients sharing one owning server. BM25 full-text search works without an embedding model; add FastEmbed for hybrid semantic search to retrieve related descriptions across code, documents and saved memory. Embeddings and remote serving are optional.
| Bring to the graph | Use it for |
|---|---|
| Code structure | Definitions, signatures, line ranges and supported relationships |
| Documents | Source text, sections and links to code |
| Agent memory | Decisions, tasks and findings with their reasons and sources |
The graph can only retrieve what was indexed or saved. Your harness chooses when to read or write, and may send retrieved context to its model provider.
Which version do these docs describe?
This guide describes grag 0.13.0 and deploys from main.
Features still in development are marked on the relevant page. This is one
maintained guide rather than a separate copy for each release.
For an additional generated architecture view, visit DeepWiki. Check its index date and source commit; its refresh schedule is independent of grag releases.