LIMITATIONS — agent-memory-manager

Everything simulated, mocked, in-process, skipped, or static-only is disclosed here. Each item is a DISCLOSED_SEAM.

DISCLOSED_SEAM: lexical (not learned-semantic) default embeddings

The default HashingEmbeddingProvider is a deterministic feature-hashing model. Cosine similarity tracks token overlap, not learned semantics. It runs offline and is reproducible, which is why it backs the verification benchmark. A hosted semantic embedding model can be plugged in via RemoteEmbeddingProvider, but is not exercised here. Consequently, the reported retrieval precision is against a synthetic, lexically-separable answer key and should not be read as semantic-retrieval quality on real text.

DISCLOSED_SEAM: extractive (not abstractive) default summarizer

ExtractiveSummarizer selects high-frequency sentences; it does not paraphrase. ClaudeSummarizer hooks Anthropic's Messages API behind the same interface but requires a network and ANTHROPIC_API_KEY and is not exercised by the verification suite.

DISCLOSED_SEAM: Postgres + pgvector runs in-process (PGlite)

Verification uses a real Postgres engine + real pgvector via PGlite (WASM), in-process. The identical SQL/pgvector code path runs against an external Postgres server through node-postgres (pg), but that external-server path, including durability across a server restart, is not exercised here.

DISCLOSED_SEAM: fleet sync is in-process

Fleet replication is verified over an in-process InProcessSyncBus (EventEmitter). A distributed broker (Redis/NATS/Kafka) implementing the same SyncBus interface — including network partitions, ordering, and delivery guarantees — is not exercised here.

DISCLOSED_SEAM: synthetic benchmark input

The retrieval benchmark corpus is generated by a seeded PRNG with a known topic answer key. It is not real agent traffic. Absolute precision on real workloads will differ. This is the blocking gap for an official benchmark and for PRODUCTION_VALIDATED.

Not addressed (out of scope for this outcome)

• No authentication, authorization, multi-tenant isolation, or encryption of

memory contents is implemented; namespace is a logical scope, not a security boundary.

• No throughput/latency benchmark at fleet scale is claimed.
• Conflict resolution for fleet sync is last-writer-wins by updatedAt; no

vector-clock / CRDT semantics are implemented.