Screenshots
Light and dark, transitioning as you scroll
Keep scrolling and the screenshots smoothly transition from light to dark theme.
ME MEMORY is an intelligent memory system with proactive memory extraction, hybrid semantic retrieval, confidence accumulation, and a layered memory architecture. It automatically distills preferences and identity, visualizes memory decay, and gives large models truly persistent memory.
Automatically extract, retrieve, consolidate, and forget from conversations — turning scattered information into trustworthy long-term knowledge.
After each turn, the content is automatically analyzed and the LLM extracts preferences, identity, and key events worth remembering — then files them into layers.
BM25 + rule-based scoring for coarse recall, then Embedding cosine similarity for fine ranking — balancing recall and precision with instant local responses.
Repeated mentions raise confidence; negation reversals trigger conflict detection, downgrade old memories, and hand over to new ones.
Recursive chunking and summarization at 500 chars per block with 80-char overlap, up to two levels deep — compressing 2000 chars into searchable points.
Core memories are always injected; search results get dynamic word counts by rank; the context window is split across System / Memory / History.
A dual-timeline model that combines storage decay, retrieval activity, frequency, and time distribution to measure each memory's true value.
Memories are promoted by value: Working → Short-term → Long-term → Core. Core memories never fade; short-term decays fast. Consolidated every 60 seconds.
A force-directed layout reveals relationships in real time: similar memories attract, layers are color-coded. Drag, zoom, and hover to inspect.
Keep scrolling and the screenshots smoothly transition from light to dark theme.
No manual upkeep — the system runs the extract, retrieve, and inject loop in the background.
Chat as usual. When you mention preferences, identity, or key events, the system replies normally.
After each turn, the extraction LLM parses the conversation, creating new memories or matching by confidence (conflict detection + reinforcement).
Before answering user-related questions, it semantically retrieves and injects layered context, then saves the conversation snapshot.
Built on mainstream frontend tech, with the Embedding model running locally — no backend server needed.
No cloning required — the local model runs offline and your data stays entirely in your environment.
Run it, start chatting — your memory grows on its own.