dgui-hypermem
Self-hosted memory MCP server on Cloudflare Workers where a Jev layer judges memory type, salience and durability and re-ranks recall.
Overview
DGUI-HyperMem is a self-hosted long-term memory MCP server for AI agents that runs entirely on Cloudflare Workers. When storing, a Jev reasoning layer assigns each memory a type, a salience score and a durability judgment, and non-durable chatter is dropped. Recall fuses Vectorize nearest-neighbour search with D1 full-text BM25 by reciprocal rank, then Jev re-ranks candidates so the best memory wins even when wording differs. Near-duplicate or contradicting memories are superseded automatically.
It exposes Streamable HTTP MCP tools (add, search, list, profile, forget and others) and a REST mirror, and works with clients such as OpenCode. Every Jev decision is logged and flushed to a Hugging Face dataset, building a training corpus for the reasoning layer over time. A REST mirror of the MCP tools is also exposed.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What type of memory is this? | — |
| Score | How salient is this memory? | — |
| Yes / no | Is this memory durable enough to keep? | yes · no |
| Choice | Which recalled candidate best matches the query?(inferred) | — |
How it works
- Agent adds a memory via MCP
- Jev judges type, salience and durability; non-durable items are dropped
- Search fuses vector and full-text candidates, then Jev re-ranks
- Decisions are logged to a Hugging Face dataset
Strengths
- Runs serverless on Cloudflare
- Hybrid retrieval with decision-model re-ranking
- Automatic supersession of duplicates
Caveats
- Logging decisions to a public dataset may not suit private data
- Tied to Cloudflare services
Getting started
Deploy the Worker to Cloudflare as described in the README and connect an MCP client to /mcp.At a glance
- Type
- MCP server
- Status
- prototype
- Language
- TypeScript
- Runtime
- Cloudflare Workers MCP server
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
DGUI-HyperMem (DeckerGUI HyperMemory) - self-hosted hybrid memory MCP server on Cloudflare Workers with a JEV (Choice/Noul/Score) reasoning layer and a HuggingFace training-brain flywheel.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide?
Memory type, salience and durability on store, and re-ranking on recall.
Where does it run?
Entirely on Cloudflare Workers with D1 and Vectorize.
What happens to Jev decisions?
They are logged to a Hugging Face dataset as training data.
Related projects
Local long-term memory for coding agents with an optional Jev reranker that orders recalled memories.
ProjectproductionTypeScriptMulti-user MCP 'agents' cloud' on Cloudflare Workers with an optional search stage that reranks candidates with Workers AI Jev scores.
MCP serverbetaTypeScriptAgent skill where Ego Lite drives the browser and Jev resolves semantic choices: targets, filtering, ranking and evidence checks.
ProjectbetaMarkdown + TypeScriptMCP tool that asks Jev which files, web pages or snippets are relevant to a query, so the agent reads selectively.
MCP serverbetaTypeScriptMCP server with record and recall tools: facts are logged during a session and Jev keeps only those relevant to the next task.
MCP serverbetaTypeScriptFilters an assistant's memories by relevance, not resemblance: one Jev yes/no per memory in a single request, kept above a threshold.
ProjectbetaPython