dgui-hypermem

September 19, 2026 · View on GitHub

DGUI-HyperMem (DeckerGUI HyperMemory) — a self-hosted, hybrid long-term memory MCP server for AI agents, running entirely on Cloudflare Workers. Retrieval fuses vector search with full-text search, then a JEV reasoning layer (Choice / Noul / Score) re-ranks and curates what gets remembered.

A DeckerGUI project.

What it does

Stores durable memories — an LLM/JEV layer assigns each memory a type, a salience score and a durability judgement; non-durable chatter is dropped.
Recalls hybrid — Vectorize ANN + D1 FTS5 BM25 candidates, fused by reciprocal rank, then re-ranked by JEV so the best memory wins even when wording differs.
Self-maintains — near-duplicate/contradicting memories are superseded automatically.
Gets better with use — every JEV decision is logged and flushed to a HuggingFace dataset, building a training corpus for the reasoning layer.

                 ┌───────────────────────────── Cloudflare Worker ─────────────────────────────┐
   MCP client    │                                                                             │
 (opencode, etc.)│   /mcp  ─ Streamable HTTP MCP      add · search · list · profile · forget    │
      ───────────┼─► /api/* ─ REST mirror              · help · sync_jev_dataset · jev_queue_stats│
                 │                                            │                                │
                 │                          ┌─────────────────┴─────────────────┐              │
                 │                          │            JEV layer              │              │
                 │                          │  Choice · Noul · Score (systemone)│              │
                 │                          └─────────────────┬─────────────────┘              │
                 │            ┌───────────────┬───────────────┴───────────┐                    │
                 │       D1 (SQLite)   Vectorize (768d)            Workers AI               │
                 │   memories + FTS5   embeddings                fallback model            │
                 │            └───────────────┴───────────────┬───────────┘                    │
                 │                                    jev_examples queue                        │
                 │                                            │ hourly cron                    │
                 └────────────────────────────────────────────┼───────────────────────────────┘
                                                              ▼
                                        🤗 ctaxnagomi/DGUI_HYPERMEM-JEV  (training brain)

Stack

Layer	Choice
Runtime	Cloudflare Workers (ES modules, `nodejs_compat`)
Protocol	MCP Streamable HTTP (stateless, JSON responses) via `@modelcontextprotocol/sdk`
Database	Cloudflare D1 — `memories`, external-content FTS5 `memories_fts`, `events`, `jev_examples`
Vectors	Cloudflare Vectorize — `@cf/baai/bge-base-en-v1.5`, 768d cosine
Reasoning	TypeSafe AI Jev / System One (`choice` / `noul` / `score`), Workers AI fallback

Endpoints

Route	Method	Purpose
`/mcp`	POST	MCP (Streamable HTTP). Requires auth.
`/api/add` `/api/search` `/api/list` `/api/profile` `/api/forget`	POST	REST mirror of the tools
`/api/sync_jev`	POST	Flush queued JEV examples to the dataset now
`/api/jev_queue_stats`	GET	Queue status (pending / uploaded / error)
`/health`	GET	Unauthenticated status

MCP tools: add, search, list, profile, forget, help, sync_jev_dataset, jev_queue_stats.

Auth accepts Authorization: Bearer <MCP_TOKEN>, x-api-key: <MCP_TOKEN>, or ?token=.

Deploy

npm install

# 1. Create resources (once)
npx wrangler d1 create dgui-hypermem
npx wrangler vectorize create dgui-hypermem --dimensions=768 --metric=cosine

# 2. Point wrangler.jsonc at your D1 id, then apply migrations
npx wrangler d1 migrations apply dgui-hypermem --remote

# 3. Secrets
npx wrangler secret put MCP_TOKEN      # bearer token clients must present
npx wrangler secret put TYPESAFE_API_KEY   # optional: enables the TypeSafe JEV backend
npx wrangler secret put HF_TOKEN       # optional: enables dataset sync

# 4. Ship
npx wrangler deploy

Configuration (`wrangler.jsonc` vars)

Var	Default	Meaning
`JEV_MODE`	`auto`	`auto` \| `typesafe` \| `workers-ai` \| `off`
`JEV_MODEL`	`jev-latest`	TypeSafe model alias
`FALLBACK_MODEL`	`@cf/meta/llama-3.1-8b-fast-v2`	Workers AI fallback
`EMBED_MODEL`	`@cf/baai/bge-base-en-v1.5`	Embedding model
`DEFAULT_SCOPE`	`default`	Default memory namespace
`HF_DATASET`	`ctaxnagomi/DGUI_HYPERMEM-JEV`	Training-brain dataset

Wire it into an MCP client

{
  "mcp": {
    "dgui-hypermem": {
      "type": "remote",
      "url": "https://dgui-hypermem.<your-subdomain>.workers.dev/mcp",
      "enabled": true,
      "headers": { "Authorization": "Bearer {env:DGUI_HYPERMEM_TOKEN}" }
    }
  }
}

The training brain

Every JEV decision is recorded as an instruction row — the exact state and typed questions that were sent, and the answers that came back — in the jev_examples table. An hourly cron (17 * * * *) and the sync_jev_dataset tool append pending rows to ctaxnagomi/DGUI_HYPERMEM-JEV, updating train.jsonl and running totals in metadata.json.

`use_case`	`instruct_type`	Recorded when
`analyze`	`choice_noul_score`	A memory is stored/typed
`rerank`	`noul`	Recall re-ranks candidates
`supersede`	`noul`	A contradiction check runs

Layout

src/
  index.ts      MCP server, REST routes, scheduled handler
  store.ts      add / search / list / profile / forget (hybrid retrieval)
  jev.ts        JEV layer: analyze, rerank, supersede (TypeSafe + Workers AI)
  dataset.ts    jev_examples queue + HuggingFace flush
  types.ts      shared types
  util.ts       ids, hashing, base64, timing-safe compare
migrations/     0001_init.sql, 0002_jev_examples.sql
wrangler.jsonc  bindings, vars, cron

Credits

DGUI-HyperMem is a DeckerGUI project.

Who	Contribution	Link
TypeSafe AI	Jev — the first System One model — and the Choice / Noul / Score primitives the reasoning layer is built on.	https://typesafe.ai · https://docs.typesafe.ai
DeckerGUI	Design, implementation and operation.	https://deckergui.my
KrackedDevs	Community credit and support.	KrackedDevs
CTECX	Knowledge / corpus partner.	CTECX

Jev, System One, and the Choice / Noul / Score primitives are TypeSafe AI's and are used under their MIT-licensed public documentation (https://docs.typesafe.ai).

Contributing

Contributions are welcome. See CONTRIBUTING.md; contributors are listed in CONTRIBUTORS.md.