Issues

Retrieval quality: hardcoded _QUERY_EXPANSIONS overfits the benchmark corpus; substring token matching causes false positives

#27

· krishna3554 opened

on Aug 26, 2026

Security hardening follow-up: no rate limiting or brute-force protection on bearer auth

#26

· krishna3554 opened

on Aug 26, 2026

DX: SDK has no retry despite idempotency-safe writes; langgraph example targets nonexistent api.dma.dev; MCP server never closes its client

#25

· krishna3554 opened

on Aug 26, 2026

Recall API semantics to pin down: empty types[] silently means 'all types'; scores are per-query normalized

#24

· krishna3554 opened

on Aug 26, 2026

Correctness: recall applies the Python precision filter after SQL LIMIT, so it can return fewer than limit results

#23

· krishna3554 opened

on Aug 26, 2026

Reliability: idempotency_keys table grows unboundedly (no TTL or cleanup)

#22

· krishna3554 opened

on Aug 26, 2026

Contract drift: error responses are not RFC 9457 problem+json as declared in the OpenAPI spec

#21

· krishna3554 opened

on Aug 26, 2026

Security: unbounded metadata size allows memory-exhaustion DoS (5.2 MB accepted per memory)

#20

· krishna3554 opened

on Aug 26, 2026

Performance: semantic duplicate detection scans every stored semantic memory inside the write transaction

#19

· krishna3554 opened

on Aug 26, 2026

Code review: 3 reproduced correctness bugs (500 on naive expires_at, timezone-offset expiry filtering, 'current/latest' recall filter) + hardening gaps

#9

· krishna3554 opened

on Aug 26, 2026

Navigation Menu

All issues

Retrieval quality: hardcoded _QUERY_EXPANSIONS overfits the benchmark corpus; substring token matching causes false positives

Security hardening follow-up: no rate limiting or brute-force protection on bearer auth

DX: SDK has no retry despite idempotency-safe writes; langgraph example targets nonexistent api.dma.dev; MCP server never closes its client

Recall API semantics to pin down: empty types[] silently means 'all types'; scores are per-query normalized

Correctness: recall applies the Python precision filter after SQL LIMIT, so it can return fewer than limit results

Reliability: idempotency_keys table grows unboundedly (no TTL or cleanup)

Contract drift: error responses are not RFC 9457 problem+json as declared in the OpenAPI spec

Security: unbounded metadata size allows memory-exhaustion DoS (5.2 MB accepted per memory)

Performance: semantic duplicate detection scans every stored semantic memory inside the write transaction

Code review: 3 reproduced correctness bugs (500 on naive expires_at, timezone-offset expiry filtering, 'current/latest' recall filter) + hardening gaps

Search results