Skip to main content
You do not have to start over. If your agents already keep memory in other tools, Orchestra can read it, with your permission, tool by tool. Your tools keep working exactly as before: Orchestra only reads, and never writes into their files.

When you run orchestra enable

orchestra enable offers the memory this repository already has: its CLAUDE.md, AGENTS.md and GEMINI.md files (and other rules files in it), and Claude Code auto memory for this repository. It lists them once, on one line, with how many memories each holds, and asks one question for all of them. Memory from your other projects, and your own files for every repo, is counted on one line and not read:
To be asked about every tool and project found on this machine, one question per tool, run orchestra enable --local --all-projects, or orchestra sources connect at any time. Nothing is read before you answer. Run without a terminal (in a script or by an agent), it lists what it found, reads nothing, and tells you to run orchestra sources connect. Orchestra’s own capture of your sessions keeps running either way. If one place cannot be read, the others are still read, and the one that failed is named with the reason. Everything below works the same any time later.

See what is on your machine

This lists every supported source it finds, with how many memories each holds and when it last changed. Nothing is read or sent until you say yes.

Choose what Orchestra reads

For each tool you choose one of three things (a tool with memory in several places, such as Claude Code auto memory for several projects, is one question): When it is done you see how many memories were read, from which tools, and one example with where it came from. To connect without being asked, name the tools: orchestra sources connect cursor --yes.

What it reads

Files checked into a repository are shared with your workspace, like the repository’s sessions. Files in your home folder, auto memory, CLAUDE.local.md, claude-mem, agentmemory and Codex memory are personal: only you can recall them. claude-mem observations it marked sensitive are never read. Codex’s raw inputs (raw_memories.md, rollout_summaries/) are not read either: they summarize sessions Orchestra already captures, and Codex folds them into the files above.

How it shows up

Every memory read from another tool keeps where it came from. When you or an agent recall it, the citation names the tool and the exact place, for example CLAUDE.md:12 (Testing). If the same fact is written in two tools, recall shows it once and cites both. Secrets are removed before anything is sent, the same way they are for your sessions.

Keeping it in step

Connected sources are read again in the background when an agent session starts. To read them now:
When you edit a line in the tool, the memory changes with it. When you delete one, it stops being current. orchestra doctor shows when each source was last read, whether anything went wrong, and whether the tool’s version is newer or older than the versions Orchestra was tested with.

Stop reading a source

Its memories stop being current in Orchestra. Its files are not touched.

Move off a memory tool

If you are leaving a tool such as agentmemory or claude-mem, import its memories once instead of keeping it connected:
It shows what it found and asks before reading anything. Each memory keeps its citation to the tool and the record it came from, and stays in Orchestra after you remove the tool. Running it again adds only what is new: nothing is duplicated, and nothing you deleted in the tool since is removed here.