Limitations¶
MemHouse 0.3.0 is a community beta. These capabilities are unavailable.
The machine-readable inventory is
specs/eval/surface-contract-inventory.json,
which marks these surfaces unavailable. Release checks reject contrary claims.
Not implemented¶
Generated AshJsonApi OpenAPI. There is no published machine-readable description of the HTTP surface. Use the HTTP API reference.
Complete generated TypeScript and Python clients. The modules under sdk/
are transport-neutral skill-readiness helpers — no HTTP client, no
authentication, no retries, no pagination, and nothing published to npm or
PyPI. See SDK helpers.
The OpenAI-compatible and Anthropic-compatible gateway proxy. You cannot point an existing OpenAI client at MemHouse and have memory injected transparently.
The full grounded ask dialectic loop with in-loop citation verification.
ask retrieves, answers, cites, and abstains — but the multi-turn dialectic
with citation verification inside the loop is not there.
Connector administration and Account archive administration UIs. Both are driven from the command line. See Mix tasks. The web console shows connector status and sync history but cannot create, edit, or trigger one.
What the web console does not do¶
It does not edit gate rules or retrieval tunings. Both change behaviour for everyone in a scope, so the operations page shows what is in force and nothing more; changing them is a database or code-level act today.
It does not administer peers, roles, or API keys. Role grants are listed on the scope directory, but granting, revoking, and provisioning are Mix tasks.
It has no bulk export. Panels are capped at 50 rows and the explorer pages 25 at a time. A complete copy of an Account is the portability archive — see Export and import.
It never shows the entity cache. Entity and mention rows span every scope that mentioned a name, so no canonical name, alias, or entity identifier appears in the console, including in the graph. That is a permanent design rule, not a missing feature.
A graph hub may still be named. The name is a wording taken from the statements in that hub's own scope, which the same panel already shows you, so it crosses no boundary. Hubs whose statements resolved to more than one referent stay unnamed. Account admins see aggregate resolution quality signals, and statement detail links only co-mentioned statements the reader may already read.
It shows no embedding vectors or chunk contents, only their counts and identities. They are rebuildable caches.
Readiness checks are for yourself only. Checking another peer's readiness in the browser would disclose which knowledge exists about them in scopes you may not hold; the JSON surface is where an agent checks its own peer.
Evaluation evidence is deliberately small¶
Upstream-scale scores and independent live-model judge evidence do not exist. Committed fixtures are smoke-scale; recorded reports are historical baselines, not current performance claims.
Published quality numbers require application version, retrieval profile and version, four model-role versions, dataset id/SHA-256/split, deadline, date, judge, strategy override, and run limits.
Operational gaps worth planning around¶
Bootstrapping a packaged release goes through bin/memhouse rpc rather
than a first-run wizard.
Ortex embedding artefacts are operator-supplied. The embedder downloads nothing. Semantic retrieval requires you to provide the ONNX model and tokenizer.
No job pruning. Nothing deletes rows from the Oban jobs table; plan for that growth.
Single Account. The community build serves one Account. Multi-Account isolation is an enterprise concern.
What is not a limitation¶
These are deliberate designs, not gaps:
| Behaviour | Why it is intentional |
|---|---|
ask abstains |
An answer invented from an empty candidate set is worse than silence |
| Extracted knowledge is not immediately visible to everyone | Blast radius scales the bar; that is the product |
| Agents cannot approve their own submissions | The gates exist precisely to prevent this |
| Entity data is unreachable through every API | A resolution error must never cross a scope or Account boundary |
| No embedded alternative database engine | Every lane runs PostgreSQL, so no lane tests software nobody ships |
Where the outstanding work is tracked¶
Acceptance criteria live in
specs/roadmap/beta-roadmap.md.
Current behavior and verification live in
specs/implementation-status.md.