Building TenseMem around a simple failure mode: agents don t forget facts - they forget which ones still apply (prefs flip, corrections get buried, retrieval serves two current truths).

Curious what you ve seen in the wild:

• stale preferences winning

• "I already told you that" loops

• tool/system text contaminating user memory

• something else?

What broke for you?

Our bet: the model proposes typed ops; deterministic rules apply them (cardinality, freshness, corrections that stick).

If you ve built agent memory - where do you draw that line? Pure prompt/tools, or engine-side resolution?

Alpha open source. Default is one SQLite file, no server.

pip install tensemem

Looking for builders who ve fought conflicting prefs / stale context. Roast the API, docs, or Studio especially anything confusing in the first 10 minutes.

1. Perfect recall of everything

2. One clear current answer per slot

3. Freshness ( is this applied yet? )

4. Corrections that survive rebuilds

We're biased to 2 4 - curious if that matches how you ship.

Happy to dig into architecture, SQLite single-writer, Studio, or "why not just embeddings." Ask anything - including "this is overkill."