txtai had the cleanest supersession semantics of 4 memory substrates in a conformance suite I built

[upcomingsimplecoder]

Hi David,
I built a 17-case conformance suite for AI memory trust primitives
(signed deletion receipts, default-DENY policy contracts, bi-temporal
validity) and ran four upstream substrates against it via thin HTTP
adapters.. no shimming, no simulation, just calls to the real packages.

Results:

graphiti (Neo4j 5)   11/17
Letta v0.16.7         9/17
txtai 9.7.0           9/17
mem0 2.0.1            8/17

Two notes specifically about txtai:

1. Embeddings.upsert(id, ...) gave the cleanest supersession
   behaviour of any substrate I tested. Same id -> replaced row,
   exactly what the conformance case expects. mem0 and Letta both
   needed adapter logic to map "kind+key" to their internal model;
   txtai just worked.

2. The only thing standing between txtai and a higher score on the
   optional bi-temporal cases is surfacing valid_at / invalid_at
   on search() results. The data already lives in the metadata
   dict, it's just not first-class on the result row.

Adapter (~150 lines, deterministic id from sha256(kind+key), HTTP
wrapper around an Embeddings instance per tenant):
https://github.com/upcomingsimplecoder/mosaic/blob/main/adapters/txtai_real.py

Full suite + reproduction:
https://github.com/upcomingsimplecoder/mosaic

Happy to send a PR exposing optional first-class temporal fields if
that direction is interesting. Either way — txtai was a pleasure to
integrate with, wanted to say so.

Thanks,
Jed

[davidmezzetti]

Hello - Thank you for this write up and sharing this information. I really appreciate it.

I'm not quite following. So you're suggesting we add valid_at and invalid_at fields to embeddings.search?