Aru and Cece have released "Lossless-memory," an open-source project designed to provide long-term memory for personal AI assistants without using summarization. Unlike many existing systems that compress conversations into compact notes or rely on vector embeddings for similarity-based retrieval, this system preserves every line of dialogue with an ISO-8601 timestamp.
Lossless-memory project enables personal AI memory without summarization
The system operates by converting conversation turns into fixed seven-field records, which are then appended to a per-day JSONL file. These raw logs serve as the single source of truth. From these logs, the system builds three types of indexes: an exact SQLite FTS5 index for keyword and timestamp searches, a vector index using sqlite-vec for semantic meaning, and "LLL" (a tiny index of topic markers) to track when conversation subjects shift.
By using the raw logs as the primary axis, the AI can retrieve exact historical statements—such as specific decisions made at a particular time—rather than providing paraphrased summaries. This approach deliberately trades increased disk space for the ability to maintain the original context and exact wording of past interactions. The implementation has been running daily for a single user since July 2026, utilizing raw logs dating back to June 2026.
Sources
- Show HN: Lossless-memory – a personal AI memory that never summarizes （Hacker News Frontpage, 2026-09-21）