Your AI Can Remember Everything, Right on Your Computer
aru-labs has published lossless-memory, a memory system for personal AI that keeps full conversations on your own computer with no summarizing step. aru-labs/lossless-memory
It is built for a simple setup. One person and one AI on one machine. No server and no cloud.
It stores chats in two linked forms. Daily text log files in JSONL format, where each line is one entry, are paired with SQLite, a simple database that lives in a single file. It uses FTS5 for exact word search and sqlite-vec for search by meaning, which can find an old chat by idea even if the words differ. Both searches run inside that same local file.
Every chat turn is saved the same way. It becomes a record with seven fields: ts, actor, role, type, text, model and session, added to that day's log file. Every record has a timestamp. That timestamp sets the order for storing and finding entries.
Search is built on time. The project calls this the Temporal Backbone. A small helper index called LLL is given to the AI on every turn so it knows where it stands. The full log stays the source of truth, and the indexes are ways to find things in it.
It has run every day since July 2026 as the memory for one person's AI assistant, with raw logs going back to June 2026. aru-labs/lossless-memory
The broader context here is a choice between keeping everything and keeping things tidy. Most AI memory systems shorten, compress or fade out old chats. That limits storage growth, controls embedding costs and keeps context use predictable. Lossless-memory does the opposite. It keeps every line word for word and decides what matters only when you ask.
In my view, the bet is that search by time, exact words and meaning can stay accurate enough without a middle step that rewrites history.
In my view, skipping summaries changes what can go wrong in ways users will recognize. Summaries drift. They mix up who said what, lose the order of events and drop small details that later matter. It is like keeping a full recording instead of relying on notes. Keeping raw text avoids those errors. The trade is that search must work harder, and the LLL snippet given to the AI each turn must stay small while still pointing into a file that grows daily.
Looking at what this means for builders, staying on one machine shapes the whole design. Local log files plus SQLite remove syncing, multi-user support and remote search. Tracking stays simple. Each turn can be traced by session, actor, role and model, ordered by ts. The limit is clear too. It does not handle sharing, moving between machines or multiple writers at once. For a personal assistant where steady memory matters more than teamwork, that may be a fair trade. Time is the main key, word and meaning search are secondary paths, and long-term context stays on your computer rather than with a remote service that decides what to keep.
