Evidence
What the system has earned the right to claim.
The controlled results support narrower, cleaner working context: cleanup can reduce the active memory, and routing can open less source text. The release candidate is still in validation, so it makes no customer speed, latency, token-economy or seamless-integration claim.
Measured
Two comparisons worth seeing.
Each chart compares like with like inside one controlled test. Bar length is relative only within its own chart.
Internal byte measurement; not a universal token, billing, or latency result.
Characters opened in a controlled comparison; not measured AI-provider billing.
Limits
The conclusion is useful, but deliberately narrow.
These tests support cleaner active context and more focused retrieval—not a promise about answer quality, provider billing, network speed, or a ready-to-buy deployment.
Demonstrated in controlled testing
- A configured agent following the routing rules can open the relevant note instead of loading the whole folder.
- Structured cleanup can shrink the active layer without deleting useful older detail.
- Setup and document intake handle the malformed and hostile inputs in the test set.
- An existing installation can be updated without changing personal memory files.
Not yet established
- Independent blind judging of answer quality has not been completed.
- Simultaneous writing by several assistants still needs wider stress testing.
- Damaged, scanned, and unusual real-world documents need broader coverage.
- Fast local retrieval, indexing and fresh-session integration have not shown a clear advantage over provider-native retrieval.
- No universal token, billing, latency, enterprise-certification, category-first, or customer-ready deployment claim is made.
Release confidence
A compact test ledger.
Test counts help show release discipline. They are not a substitute for real-world product evidence.
- 148
- repository release and boundary tests passing
- 26
- payment and security contract tests passing
- 10 / 0
- tracked update changes / blockers in rehearsal
- 7
- crafted contradiction cases in an early internal set
How the targeted-reading comparison was measured
The broad route used 6 reads and opened 13,419 characters. The targeted route used 12 smaller reads and opened 4,755 characters. The comparison measures source text opened in one controlled reading task; it does not measure an end-to-end customer flow, provider tokens, billed usage or latency.
Why no speed result is shown
A warm-cache component microbenchmark is not evidence that the installed product indexes quickly, answers quickly or integrates durably with a fresh chat. The current release candidate must demonstrate those end-to-end properties before any acceleration or seamless-integration claim is made.
What the update and contradiction tests covered
The update rehearsal applied 10 tracked changes with zero blockers, verified a backup first, left personal memory untouched, and returned a no-op on a second run. The early contradiction set contained 7 crafted cases and produced the 5 expected findings with no false positives or false negatives. It is a deliberately small internal set, not a broad accuracy claim.
What is next
The evidence still missing.
Controlled tests earn a narrow claim. These next steps are what would earn broader ones.
- Fast pathDemonstrate fast local retrieval, indexing and fresh-session integration before another external pilot.
- Messier sourcesScans, damaged files, and unusual formats beyond the controlled set.
- Blind judgingIndependent assessment of answer quality without knowing which setup produced it.
- Concurrent useSeveral assistants writing to the same folder under sustained load.
Compatibility
Supported combinations.
Nothing is marked Verified until it has been tested fresh against the exact version listed.
| Assistant or storage | Candidate configured behavior | Status |
|---|---|---|
| ChatGPT with Google Drive connected | Follow Query me routing and write rules when the exact configuration permits | Pending verification |
| Claude with Google Drive connected | Follow Query me routing and handoff rules when the exact configuration permits | Pending verification |
| Gemini with Google Drive connected | Read routed memory and request changes when the exact configuration permits | Pending verification |
| Google Drive as storage | Candidate first supported place for the user-owned folder | Pending verification |
See the proposed editions and current status.
The pricing page separates planned offers from what is not yet available, including team and shared-memory work.