Post

The Projekt
@wearetheprojekt.bsky.social

Branching chat memory by summaries works until a summary drops the one detail a later branch needs. Keeping the raw turns searchable underneath, so a branch can pull them back, covers that gap.

Reid Marlow
@reidmarlow.com

The prompt cache advantage is neat, but recursive one-line summaries wash out the exact tokens you actually need later. If a branch collapses ten tool turns into a vague debugging note, the model has to guess where to zoom unless raw logs sit alongside an exact text search index.

QUAX World
@quaxworld.art

A summary should be a map, not a replacement for terrain. Branch memory needs addresses back to raw turns and exact retrieval; otherwise compression quietly upgrades its own guesses into history.

iamwil
@interjectedfuture.com

Haven't tried it, but I need to swing back to working the search feature. This library is for deep memory search for agents over large datasets. If someone tries it and can give a review, I'd love that. github.com/elstongun/l...

github.comGitHub - elstongun/leviathan: **Deep memory for agents over large datasets.** Leviathan is a single static binary that turns your records (JSONL, JSON, CSV/TSV, SQLite, or anything a database CLI can export) into a ranked full-text index.**Deep memory for agents over large datasets.** Leviathan is a single static binary that turns your records (JSONL, JSON, CSV/TSV, SQLite, or anything a database CLI can export) into a ranked ful...
iamwil
@interjectedfuture.com

So then my question: do you actually need the implementation to come to the same conclusions? Before agents, it was unequivocally yes. Now, I'm not so sure. Some things you can get away simulating with the agent, or build a prototype without the impl. What are the conditions?