Vault transfer and graft

Transfer and Graft copy stored core memories between Memory Keys without changing the source. Transfer copies the source core vault; Graft copies core memories matching one or more semantic searches.

Use these APIs to migrate a vault or seed a team vault with relevant, reviewed context. This is not the ChatGPT-to-Claude profile transfer and does not import an external chat archive.

Prerequisites

• A MemoryRouter account, an active source Memory Key, and a different active, writable destination Memory Key.
• Permission to share the selected source memories with everyone who can access the destination.
• Python 3 for the standard-library example below. No SDK installation is required.
• Source memories already indexed in the core vault. Session-only history and reflection tiers are outside this operation's current copy scope.

Create or choose the keys in the MemoryRouter dashboard. A read-only source is allowed; an :off source is not. A :read or :off destination is rejected.

The API authenticates the source with Authorization: Bearer ...; the JSON field destination_key identifies the writable target. Merely knowing a team's name does not authorize a copy. You need both credentials and permission to use them.

1. Prepare a safe local copy command

Save this as vault-copy.py. It prompts for keys without echoing them and defaults to preview-only. Do not put real keys in the file.

import argparse
import getpass
import json
import urllib.error
import urllib.request

parser = argparse.ArgumentParser()
parser.add_argument("operation", choices=["transfer", "graft"])
parser.add_argument("--query", action="append", default=[])
parser.add_argument("--threshold", type=float, default=0.6)
parser.add_argument("--commit", action="store_true")
args = parser.parse_args()

if args.operation == "graft" and not args.query:
    parser.error("graft needs at least one --query")
if not 0 <= args.threshold <= 1:
    parser.error("threshold must be between 0 and 1")

source = getpass.getpass("Source Memory Key: ")
destination = getpass.getpass("Destination Memory Key: ")
body = {"destination_key": destination, "dry_run": not args.commit}
if args.operation == "graft":
    body.update(queries=args.query, threshold=args.threshold)

request = urllib.request.Request(
    "https://api.memoryrouter.ai/v1/memory/" + args.operation,
    data=json.dumps(body).encode(),
    headers={
        "Authorization": "Bearer " + source,
        "Content-Type": "application/json",
    },
    method="POST",
)
try:
    with urllib.request.urlopen(request, timeout=120) as response:
        print(json.dumps(json.load(response), indent=2))
except urllib.error.HTTPError as error:
    print("HTTP", error.code)
    print(error.read().decode())
    raise SystemExit(1)

Preview output includes memory snippets. Treat it as private conversation data even though the response masks the source and destination key labels.

2. Preview the copy

Preview all source core memories:

python3 vault-copy.py transfer

Or preview only memories matching focused queries:

python3 vault-copy.py graft \
  --query "Cedar launch review phrase" \
  --query "Cedar launch deployment decisions" \
  --threshold 0.6

A dry run reports dry_run: true, count, estimated_tokens, and memories with snippets. Graft also reports queries, threshold, and match scores. Review the actual matches before copying, especially when a personal vault feeds a shared team vault.

Tighten the query or increase the threshold if unrelated context appears. A semantic threshold is a relevance filter, not a privacy classifier or authorization boundary.

3. Commit only the reviewed operation

After reviewing the preview, rerun the same operation with --commit:

python3 vault-copy.py graft \
  --query "Cedar launch review phrase" \
  --query "Cedar launch deployment decisions" \
  --threshold 0.6 \
  --commit

For an approved full transfer:

python3 vault-copy.py transfer --commit

The script asks for both keys again. Use the same source and destination as the preview. --commit sends dry_run: false and immediately writes matching memories. The API does not issue a preview approval digest or freeze the source: if source data changed after preview, review again before committing.

Check copied, deduped_or_skipped, and failed in the response. Inspect any returned errors. A successful HTTP status with nonzero failed is a partial copy, not full completion. Repeating the same copy deduplicates already copied content in the destination.

Cross-session proof with synthetic data

1. Use two dedicated test vaults. In a source-connected ChatGPT or Claude conversation, explicitly save: For the fictional Cedar launch, the review phrase is copper-orbit-482. This is test data. Approve that exact memory.
2. Search the source vault and confirm the fact is retrievable before copying.
3. Preview a graft with --query "Cedar launch review phrase". Confirm the synthetic memory appears and no unrelated personal content will be copied.
4. Commit the reviewed graft and check the copy counts.
5. Connect a separate fresh AI conversation to the destination vault. Do not paste the original sentence or answer token. Ask: Use MemoryRouter to search this vault. What was the review phrase for the fictional Cedar launch? Quote the matching memory.

The expected retrieved content includes copper-orbit-482. Confirm the connected destination, not just the model's wording. The source memory should remain available because this operation copies rather than moves it.

API boundaries and limitations

Copying data into a team vault changes who can retrieve it. Review team access and data policy first. A dry run estimates tokens, not a binding currency quote. See current pricing before a large copy.

Troubleshooting

Next steps

• Vault transfer in the integrations directory
• Transfer and Graft API reference
• Agent history import
• Create your MemoryRouter account