Question 1

Is the Jev reranker worth $0.0004 a recall over the free cross-encoder?

Accepted Answer

On ranking, yes; on answer quality, the author's own tests say no. R@1 went from 0.41 to 0.62 on a private 300-query store with the margin holding in 20 of 20 seeds, but three graded tests showed no better answer rate than the free local cross-encoder, and that negative result is published in the same doc. What it buys is a shorter context: two Jev-ranked memories matched five cross-encoder-ranked ones.

Question 2

What exactly leaves my machine when I turn features on?

Accepted Answer

Recall makes no network call by default: search is BM25 with no model. Enabling --reranker jev sends the query and candidate memories to TypeSafe. Separately, `hippo sleep` sends memory text to Anthropic for fact extraction whenever ANTHROPIC_API_KEY is set, and it runs at the end of each Claude Code and OpenCode session plus a daily job; set extraction.enabled to false in .hippo/config.json to stop it.

Question 3

How much of the published benchmark claims held up?

Accepted Answer

Less than the badges suggest. One informal magnitude claim was retracted in a later release while the mechanism stayed shipped, decay tied with decay switched off, and sleep consolidation lowered recall in the author's tests. Whether superseding a memory helps has not been measured. The project does index preregistrations next to results, including the runs that failed.

Question 4

Will installing the package give me automatic memory capture?

Accepted Answer

No, and the README says package installation alone does not enable automatic preservation on every agent. You run `hippo init` inside a project: it creates the .hippo store, adds hooks for the agents it finds on your machine, schedules a daily run, and appends a block to the CLAUDE.md or AGENTS.md already there. Codex hooks only run once you trust them in /hooks.

Question 5

What does hippo init actually write?

Accepted Answer

It creates the project's .hippo/ store and adds a block to the CLAUDE.md or AGENTS.md already there. On your machine it adds hooks for whichever agents it finds, such as Claude Code's in ~/.claude/settings.json, plus a daily 6:15am run, and flags skip each part individually. Package installation alone does not enable automatic preservation; capture and compaction coverage depend on completing the setup and the host trust step.

Question 6

What are the requirements and what is optional?

Accepted Answer

Node.js 22.16 or newer and zero runtime dependencies. Search is BM25 with no model and no network call, and embeddings are an optional install: bring-your-own local Transformers.js, or an opt-in API embedder on OpenAI, Voyage, or Cohere. Nothing is auto-installed, which is why the default recall path stays local.

Question 7

How does the learning mechanism actually work?

Accepted Answer

Outcomes feed back into ranking. Mark a memory wrong and it drops out of the top results; keep using one and it strengthens; tag a failure with --tag error and it gets twice the half-life of an ordinary memory, so the lesson is still there when the next recall matches. On the synthetic mechanism audit those two parts are what measured as helping retrieval.

Hippo

About this project

Highlights

Quickstart

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community