# lossless-memory

A local, file-based long-term memory layer for a personal AI that never summarizes conversations, keeping every line with a timestamp and searching by time first and words second, rather than retrieving only similar-seeming chunks.

lossless-memory is a long-term memory system for a personal AI assistant built on the position that summarizing conversations or retrieving only semantically similar chunks both lose what matters most to someone who talks to the same assistant every day: what was actually said, and when. It stores raw conversation logs in full as JSONL files plus a SQLite database, using full-text search for exact matches and a vector extension for semantic search only as a last resort, with every record, utterance, action or document chunk carrying a timestamp on what it calls a temporal backbone.

A query such as a time phrase combined with a topic narrows the search range first and ranks words within that range second, returning unsummarized results in chronological order with their timestamps attached, and a small index tracks which topic the conversation is currently in so that identity and context survive a context-window compaction or a new session. It is designed for one person and one assistant running on one machine, with no server and no cloud component, and the project documents its design failures alongside what worked.

Website: https://note.com/aru_log