Question 1

Where does my memory data actually go?

Accepted Answer

Up to three places, depending on config. Memories live in your D1 database with an external-content FTS5 index, and embeddings live in Vectorize. With JEV_MODE at auto or typesafe, the memory text is also sent to TypeSafe for the typing, salience and durability judgements. If you set HF_TOKEN, each decision is additionally queued in the jev_examples table and flushed to a HuggingFace dataset — that is a deliberate training-corpus export, so leave the secret unset if you do not want it.

Question 2

Does hybrid retrieval actually beat plain vector search, or is re-ranking just overhead?

Accepted Answer

The design assumes lexical and semantic recall fail differently. Vectorize approximate-nearest-neighbour results and D1 FTS5 BM25 results are fused by reciprocal rank, and only the merged set is re-ranked by Jev, so a memory worded differently from the query can still win. The same layer runs supersede checks, so near-duplicates and contradictions are retired rather than left sitting alongside the original.

Question 3

Can I run it without a TypeSafe key?

Accepted Answer

Yes. TYPESAFE_API_KEY is optional — without it the reasoning layer falls back to a Workers AI model, and JEV_MODE accepts auto, typesafe, workers-ai or off, so you can turn re-ranking and analysis off entirely and keep the hybrid retrieval. The deploy still needs wrangler: create the D1 database and a 768-dimension cosine Vectorize index, apply the migrations, then deploy.

Question 4

How is it exposed to an agent, and how is it secured?

Accepted Answer

Over stateless MCP Streamable HTTP at /mcp, with a REST mirror at /api/add, /search, /list, /profile and /forget plus sync_jev_dataset and jev_queue_stats. The tools are add, search, list, profile, forget, help, sync_jev_dataset and jev_queue_stats. A single MCP_TOKEN is accepted as Authorization: Bearer, as x-api-key, or as a ?token= query parameter, and only the /health status route is reachable without it.

Question 5

What exactly gets exported to HuggingFace, and when?

Accepted Answer

Every Jev decision is recorded as an instruction row in the jev_examples table holding the exact state and typed questions sent plus the answers returned, tagged analyze, rerank or supersede. An hourly cron at 17 * * * * appends pending rows to the ctaxnagomi/DGUI_HYPERMEM-JEV dataset, updating train.jsonl and running totals in metadata.json, and the sync_jev_dataset tool forces the flush.

Question 6

Which models does it actually depend on?

Accepted Answer

Cloudflare's @cf/baai/bge-base-en-v1.5 produces the 768-dimensional cosine vectors in Vectorize, and the no-key fallback reasoning model is @cf/meta/llama-3.1-8b-fast-v2 on Workers AI. The TypeSafe path uses jev-latest by default and takes JEV_MODEL, with DEFAULT_SCOPE setting the memory namespace and JEV_MODE selecting auto, typesafe, workers-ai or off.

Question 7

How hard is the deploy, and what license is it?

Accepted Answer

MIT, copyright 2026 DeckerGUI, with TypeSafe credited for the primitives and the repo listing KrackedDevs and CTECX as contributors. Deploying is npm install, create a D1 database and a 768-dimension cosine Vectorize index, point wrangler.jsonc at the D1 id, apply the migrations, then set MCP_TOKEN plus the optional TYPESAFE_API_KEY and HF_TOKEN as secrets before npx wrangler deploy.

Dgui Hypermem

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community