AI · yume
In developmentYume
Personal AI still in development: layered memory, adaptive thinking, psyche stack.
- TypeScript
- Bun
- LangGraph
- Hono
- WebSocket
- SQLite
Frames
Mark only · product surface unfinished
Overview
Personal assistant monorepo on Bun: Hono HTTP/WebSocket, LangGraph in-process brain, React chat UI, and a TypeScript SDK. Three memory tiers (episodic checkpoints, vector store, structural brain markdown), adaptive think-or-direct routing, and a modular psyche stack. Still in development: no public product URL; case study only while the surface is unfinished and the repo stays private.
Problem
Personal chatbots usually drop long-horizon context, hard-code one personality, and hide extension points behind a single chat UI with no real memory architecture.
Approach
Keep the brain in-process on Bun: LangGraph for turn control, SQLite for episodic + vector memory, markdown brains for identity, Hono for HTTP/WS, and a separate React client + SDK. Route simple turns direct; escalate multi-step work through an internal thinking path.
Features
3-tier memory
Episodic chat in yume-checkpoints.sqlite, semantic recall via vector store items, structural identity in data/brains: each layer has a different write path and lifetime.
Adaptive thinking
Fast heuristics for greetings and short turns; LLM classifier on the grey zone; fixed deflects for injection and off-topic intro noise.
Psyche stack
Six optional cognitive layers injected into RP context (not shown raw to the user). Reconciliation can skip expensive layers for latency; image turns skip psyche/memory retrieval.
Transport + SDK
REST under /v1 for chat, memory, skills; WebSocket for live turns and activity signals; TypeScript client for tooling and automation.
Proof
- Built: LangGraph turn graph, SQLite checkpoints, adaptive classifier hooks, Hono transport.
- Not built for public: stable product URL, multi-tenant SaaS, polished marketing shell.
- Injection and off-topic guards sit in front of expensive thinking paths.
- Personal-server bias is intentional; horizontal scale is a non-goal for now.
Tradeoffs
- Still in development: not a public product; no marketing deploy or stable public URL yet.
- Repo is private; local/VPS ops, not multi-tenant SaaS.
- In-process LangGraph + SQLite favors a personal server over horizontal scale.
Need an AI system, a streaming chat backend, or a local-first product. Email is fastest.
Email brief