QilbeeDB 0.13.0: typed memory graphs
QilbeeDB 0.13.0 adds durable typed relations, bounded graph traversal, relation
change consumption and experimental graph-assisted retrieval. Check the version of your installation with GET /health before using these
contracts.
The API stores assertions supplied by applications and external workers. It does not generate embeddings, resolve entities automatically or infer relationships from memory text. Existing memories remain readable without relation backfill. New typed edges require explicit revision-bound assertions with provenance.
Available contracts#
| Capability | Contract | Integration responsibility |
|---|---|---|
| Typed assertions | Relation commands, reads and history | Supply exact endpoint revisions, provenance and stable idempotency keys; apply review authority explicitly |
| Neighborhood traversal | Typed graph API | Select roots, direction, kinds and limits; inspect coverage and treat an incomplete neighborhood as incomplete |
| Relation events | Changes, checkpoints and diagnostics | Consume both memory and relation histories; preserve sink state before progress and revalidate current eligibility |
| Graph-assisted retrieval | Graph search and immutable profiles | Opt in to a named profile and seed mode; inspect exact path evidence and coverage |
| Company administration | Retained workspace inspection and typed relation audit | Use company administrative authority to discover authorized private subjects and retained history |
The existing cosine, BM25 and hybrid routes preserve their score meanings and defaults. Graph-assisted retrieval uses a distinct endpoint and combined score; its score is neither cosine similarity nor a probability. A client must not treat a full memory scan as proof of a complete graph: candidate, traversal and embedding coverage are reported separately.
Upgrade acceptance#
Use the following sequence for an installation-specific acceptance run. Preserve existing data, credentials and immutable history throughout the procedure.
- Read
/healthand/openapi.json, record the current image identity, and verify that the intended 0.13.0 image has passed the relevant functional and security assessment. Review known image findings separately from API tests. - Take an application-consistent backup and test the new image against a stopped copy before replacing the live service. After public writes resume, do not restore an older copy without reconciling those writes.
- Verify retained record IDs, revisions and payload fingerprints, together with existing company, credential, login and scope-authority metadata. Keep those observations private; publish aggregate results only.
- In a separate synthetic scope, create two related memories and another private subject. Assert a relation and retry the identical command. Check that traversal and graph retrieval include the authorized neighbor and exclude the other subject. Verify explicit denials for unauthorized project, agent and mission scopes.
- Consume relation changes, checkpoint progress, retry the checkpoint command, and read its current state. Reject use of an old receipt as proof that the remote checkpoint has not advanced. Reuse the memory feed and current-state checks because endpoint updates and time-based expiration do not require a relation lifecycle event.
- Update an endpoint revision. Confirm that the old relation no longer contributes to current traversal or ranking, while its historical revision remains available to authorized audit readers.
- Retire owned test relations, delete owned test memories and revoke temporary credentials. Verify terminal state and denied authentication. Retain the resulting audit records and checkpoints, then confirm that preexisting data and authorities are unchanged.
Ranking remains experimental#
The frozen MuSiQue comparison used 100 reserved public-development queries, eight methods and 2,400 requests. Balanced graph retrieval changed mean nDCG@10 from 0.5311 to 0.5422 relative to hybrid v2. The paired 95% interval for the difference includes zero. Queries with no labeled support increased from 3% to 8%, and retrieval p95 increased by 26.6%.
This evidence does not qualify graph retrieval as the default. It also does not show improved agent reasoning. Evaluate the published protocol on a frozen workload, preserve regressions, and compare agent task outcomes with the same model, prompt and tools before making a broader capability claim.
Client and console boundaries#
The relation API is available to lightweight HTTP clients. A packaged relation consumer adapter and typed-relation workflows in the corporate console are separate deliverables. Updating the server does not add those features to an already deployed client. Use the HTTP reference for the released API contract, and verify the client or console version independently.