Question 1

Why does this need Jev instead of an LLM judge?

Accepted Answer

Checking all memories against each event with an LLM is too slow and costly, so products check only similar items or nothing. Jev answers one yes-or-no question quickly and cheaply, so checking everything becomes the default.

Question 2

Does it rewrite my stored memories?

Accepted Answer

It does not. The old text is marked and kept with its verdict and logged cause, and the replacement is stored verbatim. Recall then returns live memories while dead ones never reach the prompt.

Question 3

Can a prompt injection wipe my memory through this?

Accepted Answer

No. Questions, plans, and instructions such as ignore-previous-directives are gated out as hypothetical or directive evidence. Outages confirm existing facts, and unsure cases go to review rather than flipping status.

Question 4

Do I have to migrate off my current memory store?

Accepted Answer

It runs in front of Mem0, Chroma, LangGraph store, Markdown notes, pgvector, Qdrant, Letta, and Zep through adapters. The host keeps the data and invalidate writes back statuses and verdicts as metadata.

Question 5

How much does checking every memory actually cost?

Accepted Answer

One check costs $0.00006, because Jev answers the yes/no question with a probability in about 150 ms. Checking 500 memories against one event takes under a second and costs half a cent, which is why checking everything is the default rather than a premium tier. The contrast with a typical memory product is exactly that: those check the ten most similar or none, decide with an LLM or the agent itself, and either overwrite the old memory or keep it forever.

Question 6

What do the evaluation numbers cover, and how do I install it?

Accepted Answer

The eval runs on 157 labeled cases and reports 89.2% strict, 97.5% lenient, and 0 of 157 memories wrongly retired; the test suite is 695 passing tests. There is no PyPI release, so pip install invalidate does not resolve: use pip install git+https://github.com/chopratejas/invalidate, then export TYPESAFE_API_KEY or put it in a .env. From source, pip install -e ".[dev]" and pytest -q runs 519 tests with no key, and the eval harness runs dry before a live pass of 157 cases at about three seconds and two cents.

Question 7

What can I run to see it work, and what statuses will I see?

Accepted Answer

invalidate demo runs six facts against four events, invalidates three of them and takes about 200 ms per event; invalidate ui opens a local UI at http://127.0.0.1:7411; and the live playground needs no install at all. Statuses in the transcript are active, superseded, contradicted and needs_review, each with the probability of still being true, and a ls column shows id, status, p_true, kind, source, age and fact. Recall returns only the live ones.

Invalidate

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community