Hippo Memory
Hippo Memory (hippo-memory.
Overview
Hippo Memory is kitfunso's open memory layer with a hippo recall CLI, local storage, and an opt-in TypeSafe Jev reranker. Default recall stays cheap and offline-friendly; add --reranker jev when you want structured ranking judgments on candidate memories before results print.
Problem: Agent memory tools retrieve by embedding similarity alone, so the first hit is often related noise instead of the fact you actually needed.
Built for: Builders who want a local-first memory CLI and optional Jev reranking when recall quality matters more than raw speed.
First indexed on Jev Directory: 2026-09-22
Creator and team
- Name
- kitfunso
- Handle
- @kitfunso
How Jev is used
- Role in the product flow
- Optional reranker on memory recall (ranking only, not generation)
- Primitives
- ScoreNoul
- State in
- User query plus a bounded set of candidate memory rows retrieved by the local index before reranking runs.
- Decision out
- Reordered recall list where Jev scores or keeps-drop judgments elevate the best matching memory to rank one.
- Run hippo recall "<query>" to fetch local candidates
- Pass --reranker jev to enable TypeSafe Jev on the candidate set
- Jev applies ranking judgments per README schema (not summarization)
- Return reranked rows to the caller; reranker stays off unless opted in
Hippo does not ask Jev to write memories or paraphrase notes. The hot path is recall: local retrieval proposes candidates, then Jev answers structured questions about which rows best satisfy the query. That is classic ranking work for Score and Noul style probes instead of a chat completion. The reranker ships off by default so scripts and agents keep predictable latency; turn it on when your eval says first-result quality is worth a fraction of a cent. Read the published eval markdown in the repo before you quote the 0.41 to 0.62 recall-at-one jump in a slide deck; it is the author's private set, not a third-party benchmark leaderboard.
Sourced performance claims
- Private eval on three hundred queries: recall at one reportedly moves from 0.41 to 0.62 with Jev reranker enabled (docs/evals/2026-09-19-jev-reranker.md).Source: github.com/kitfunso/hippo-memory eval doc
- Documented ballpark ~$0.0004 per recall when reranker is on.Source: github.com/kitfunso/hippo-memory README
Features and stack
Features
- CLI hippo recall with optional --reranker jev
- Local memory store with ranking-only Jev path
- Published eval notes for reranker impact
- TypeSafe API key required only when reranker runs
Stack
- TypeScript
- CLI
- TypeSafe System One
- local memory index
Pricing: Local recall is yours to host; Jev reranker bills per TypeSafe usage (~$0.0004/recall per README order of magnitude).
Links
FAQ
- Does Jev rewrite stored memories?
- No. README positions Jev as a reranker on recall candidates only. Ingest and storage stay outside the Jev path unless you add your own hooks.
- How do I enable reranking?
- Use hippo recall "<query>" --reranker jev with TYPESAFE_API_KEY configured per install docs.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.