Remote + embedded audit at v1.1.0: Supermemory re-store 400s (PATCH drops containerTag), keyed-seam gaps, and verified adapter defects

[YellowSnnowmann]

Summary

An eight-lane production audit of the memory stack (run after #69/#71 merged) verified one blocker and a set of majors in the remote adapters and the embedded engine. Every finding below was traced in source; engine-API claims rest on vendor OpenAPI/SDK definitions and engine server source, not live calls. All paths are the post-#73 crates/ layout, verified present at v1.1.0 (1d501fb).

Blocker

• Supermemory upsert's update branch omits containerTag, which PATCH /v4/memories requires — crates/tinymemory-remote/src/supermemory.rs:355-366. The PATCH body is {id, newContent, metadata}; Supermemory's OpenAPI marks containerTag required on PATCH ("Required to scope the operation", 400 on missing fields) and both official SDKs enforce it as non-optional. The adapter's POST and DELETE both send the tag — PATCH alone doesn't. First store of a key works; every re-store of the same key returns 400 → MemoryError::Invalid, permanently for that key. The suite stays green because neither the supermemory double nor the conformance double enforces the field — fix should also make both doubles reject a tagless PATCH, the way sm_create polices POST.

Majors

Mem0

• delete never moved to the U2: keyed CRUD for the Mem0/Cognee adapters — investigation + design (get drops from O(N) to ≤3 requests) #69 keyed seam — crates/tinymemory-remote/src/mem0.rs:678. Resolves via entries() (whole-account walk) instead of the entry() seam upsert already uses. Self-hosted: forget dies at the ≥1000-record account ceiling even for a tiny namespace. Hosted: up to 500×200-row pages to delete one record. Fix is one line: resolve through self.entry(namespace, key). (Found independently by two audit lanes.)
• Current OSS server admin-gates the unscoped listing — the whole-store walk (GET /memories with no entity id) requires the admin role on today's mem0 server, so count()/namespace_summaries()/list(None)/forget() fail 403 with a regular key, and the adapter's "check the API key" hint sends the operator down the wrong runbook.
• Namespace-less recall always 400s against current OSS — the SelfHosted search arm sends "filters": {} when there is no namespace; the server requires at least one of user_id/agent_id/run_id. Fix direction for both: stamp agent_id on OSS writes and use it as the unscoped scope (mirrors the Cloud arm), plus an actionable 403 message.

Supermemory

• Account-wide ops page every foreign container tag — memories() lists all tags and pages each with no filter for the adapter-owned tinymemory: prefix; on a shared account, count()/list(None)/namespace_summaries()/export walk thousands of foreign pages and discard everything at decode.
• Concurrent same-key stores create permanent duplicates; delete removes only the newest and the stale twin resurrects — upsert is find-then-write with a page-walk-long race; nothing heals duplicates afterward. Fix: sweep all (namespace, key) matches on delete and on the PATCH branch so duplicates self-heal.
• The per-tag pager has no page ceiling — it trusts server-supplied totalPages unboundedly; mem0 got CLOUD_MAX_PAGES = 500 for exactly this failure class. Same ensure! shape needed.

Cognee

• First recall in a fresh namespace errors instead of answering empty — search() sends the dataset name with no find_dataset guard; real Cognee 404s ("No datasets found") when the name resolves to nothing. Every sibling op treats a missing dataset as empty; only recall diverges.
• "updatedAt": null halts the timestamp fallback chain — both backfill sites (crates/tinymemory-remote/src/cognee.rs:269 and the keyed-path copy near :311). Real Cognee serializes updatedAt: null for never-updated records (the common case); Value::get on a present-but-null key returns Some(Null), so the or_else chain never reaches createdAt and essentially every record answers an empty timestamp. Fix: find_map(|k| data.get(k).and_then(Value::as_str)) over the four candidates, plus a double row with a null updatedAt.
• Records larger than one chunk are silently invisible to recall — recall keeps only chunks whose text parses as a complete envelope; Cognee chunks uploads at the embedding model's token cap and the adapter sends no chunk_size, so an oversized record comes back as fragments that fail both parses and vanish without a warning (stored fine, visible to get/list, unfindable via recall; ~2 KB threshold on 512-token embedders).
• The multipart upload leg bypasses the typed error taxonomy — raw .send() + hand-rolled anyhow!: 400/413 isn't Invalid, 401 isn't Unauthorized, 429/503 isn't Unavailable — all flatten to Other, violating §A4. The failure suite misses it because its double fails the preceding typed GET.

Shared plumbing

• Non-2xx bodies are buffered with no size cap — the 64 MiB read_capped guard applies only to success bodies; all four error paths (crates/tinymemory-remote/src/common.rs, the response.text() sites) buffer unboundedly, of which 300 chars are ever used. A 64 KiB cap suffices.

Embedded engine (tinycortex)

• KV family: put→get misses, put→delete works — crates/tinymemory-tinycortex/src/engine/mod.rs:547 compares the RAW caller key against canonicalized stored keys over a full-namespace listing (kv_list same for prefixes), while writes canonicalize and kv_delete goes through the symmetric shim — an incoherent split. Machine-verified twice as drift, not design (the #5164 comment's own "invisible" class); trigger class is canonicalizer-rewritten keys (card-shaped digit runs — emails are deliberately excluded from the PII gate). Zero KV conformance coverage. Fix: canonicalize in kv_get/kv_list (the symmetric kv_get_namespace shim already exists), add a keyed fetch to stop the namespace walk, and add KV conformance.
• The mandatory path runs blocking SQLite inline on the async executor — and the repo documents the opposite — crates/tinymemory-core/src/store/memory_trait.rs (zero spawn_blocking under core/store; machine-verified twice). Every Memory-trait method locks the process-wide connection mutex and runs synchronous SQL inside an async fn, while the same adapter's Cargo.toml claims "every family method runs synchronous engine work off the async executor" — and the profile/episodic families that DO offload hold the same mutex from the blocking pool. Fix: move the Memory-impl bodies onto spawn_blocking like the profile family.

Minors (verified, lower blast radius)

• Taint decode is inconsistent at the fail line: absent marker → Internal (trusted, via #[default]) while malformed → ExternalSync (crates/tinymemory-api/src/types.rs:113-118); absent should fail closed too.
• Cognee: one corrupt/0-byte envelope poisons every enumeration op; count()/namespace_summaries() still fan out the 1+D+N raw walk U2: keyed CRUD for the Mem0/Cognee adapters — investigation + design (get drops from O(N) to ≤3 requests) #69 removed elsewhere; store() blocks on the full synchronous cognify pipeline inside the 60s budget (a timeout still leaves the record ingested); duplicate-name resolution assumes insertion order but the listing sorts by data_size.
• Supermemory: recall forwards limits >100 to an API whose documented max is 100 and category/session post-filters under-return silently; records are created isStatic: false, opting exact records into dynamic rewriting; 1..10,000-char content limits unenforced.
• Mem0: OSS update merges metadata so a cleared session_id stamp survives; Cloud sends "run_id": null against its own omit-don't-null rule; health probes are unauthenticated (revoked key reports Ready).
• Shared: 429 retries ignore Retry-After; valid-JSON envelope drift decodes as an empty store (absence laundering) on listing paths.
• Embedded: recall fabricates entry identity (positional id, Utc::now() timestamp, dropped session) while get/list return the truth for the same row; export re-materializes the namespace per page under the global mutex; health_check is pure path-existence against a contract that promises "able to serve" (both machine-verified).

Suggested order

1. The blocker + the two one-liners (mem0 delete-through-seam, pager ceiling) + the cognee trio (recall guard, find_map timestamps, chunk-size/keyed-fallback) — one adapter-correctness PR.
2. Error-body caps, taint fail-closed, multipart typing.
3. Embedded spawn_blocking + KV canonicalization + KV conformance.

Related

• adapters/remote: min_score is inert on Cognee, and three §A4 fixes are pinned by no discriminating test #72 tracks the untested-guard minors deferred from the Keyed CRUD: reads stop enumerating the account (#69) #71 review (per-namespace ceiling, .rev() determinism, supermemory's redundant retention) plus inert min_score — deliberately not duplicated here.
• Conformance doubles: Cognee recall shape makes upstream coverage vacuous; upload double swallows malformed multipart #67 covers the cognee conformance-double recall shape ({"results": […]} vs bare array) — the audit re-confirmed it; not duplicated here.
• No collision with Move every crate under crates/ and reach them all by feature #73 (crates/ restructure — additive, no behavior change; all paths above are the new layout) or Move the wire vocabulary into tinymemory-bus, beneath tinymemory-api #74.
• Companion product-side issue in opencompany (Brain view, multibyte snippet panics, fs-only shared-address guard, vendor pin chain) — link added after filing.

[YellowSnnowmann]

Companion product-side issue filed: tinyhumansai/opencompany#1488 (Brain view oldest-500 freeze, agent-reachable multibyte snippet panics, fs-only shared-address guard, vendor pin chain).