How to keep a long-running assistant on topic after context compaction with Lossless Memory

Who it is forA person running a long-lived AI assistant whose context window gets compacted, so it loses track of the current thread.

The problemAfter compaction or a new session, the assistant has no clear record of which subject the conversation is on.

The outcomeA small LLL topic index the model reads each turn, with priority and completion marks set by the human.

What you need first

Python with pip, and git to clone the repository
A machine you own; the README describes no server or cloud component
Conversation logs as per-day JSONL files with ts, role, and text fields, or Claude Code transcript files
Optional: a local sentence-transformers embedding model if you want semantic search as a fallback

Steps

git clone https://github.com/aru-labs/lossless-memory
cd lossless-memory
pip install -e .
Read docs/lll.md to learn the human and AI division of labor for topic markers.
Write short timestamped topic marker lines whenever the conversation moves to a new subject.
Inject the rendered LLL index into the model context at the start of each turn.

Prompt for your agent

Using the repo aru-labs/lossless-memory, read docs/lll.md and set up an LLL topic index for <assistant name>. Expected output: a short list of timestamped topic markers that can be injected each turn. Verify that the human-set priority and completion marks are not edited by the model.

Caveat. The README describes LLL as a design and gives no ready-made injection command, so you must wire it into your own assistant.