Lossless-memory Brings a Personal AI Store That Never Summarizes

A developer operating the aru-labs GitHub account has published lossless-memory, an open-source memory store for personal AI assistants that keeps entire conversation histories rather than condensing them into summaries. The repository, at github.com/aru-labs/lossless-memory, appeared under a Show HN submission that had drawn five points and no comments at the time of writing, leaving the implementation largely untested by outside users.

The project's central design choice runs against the grain of most assistant memory systems. Retrieval and summarization pipelines typically fold older exchanges into a running summary so prompts stay inside a model's context window, a trade that cuts token costs but blurs specifics. lossless-memory stores the raw exchanges instead, on the argument that storage is cheap enough and search good enough that nothing needs to be paraphrased away.

That distinction matters most over long time horizons. An assistant used for months accumulates thousands of exchanges, and a summary can smooth over the exact figure, name or instruction a user later needs back. Keeping the original text means a retrieval step has to locate the right passage at query time rather than reconstructing it from a compressed record.

Memory has become one of the busiest areas of applied AI work, with vendors selling retrieval and summarization layers that trade disk space for cheaper prompts. The pitch behind lossless-memory is that this tradeoff is no longer necessary, though the repository does not yet carry benchmarks on recall accuracy, query latency or token spend that would show whether that holds as a history grows past hundreds of thousands of tokens.

The obvious pressure point is retrieval cost. Pulling exact passages from an ever-growing archive can be slower and more expensive per query than reading a short summary, and search quality decides whether the correct memory surfaces at the right moment. Projects in this space tend to stay demos until they publish numbers on those three axes.

What to watch is whether aru-labs adds evaluation data and how the implementation holds up once developers run it against their own archives. A usable local setup could give developers a reference point for assistants that remember verbatim; without it, the summarizing incumbents keep the field to themselves.