Ramon Fabrega

lore

Claude Code writes constantly and keeps almost none of it. Three failures drove this tool.

The lore explorer: transactions, steps and instructions for one session, with a say/read/write/run timeline and per-transaction latency. — A session as a block

The lore explorer: spend by day, by model and by well, split into input, cache write, cache read and output. — Where the tokens go

The lore explorer: every session recorded against a single directory, oldest first, with models, requests and cost. — One directory, in order

Sample data — a synthetic corpus, so nothing here is anyone's real work. The tool reads your own.

Why

Evaporation

Default retention deletes transcripts after about 30 days. Check cleanupPeriodDays in your settings — by the time you go looking for something older, it is already gone.

Path-sharding

Memory and transcripts are keyed by directory. A repo, a worktree inside it, and the wider folder you sometimes work from are separate, mutually blind stores — nesting counts for nothing. What a session learns in one place is invisible everywhere else.

No compounding

Decisions and corrections sit inert in JSONL. Nothing accumulates. Every session starts over.

What it is

A CLI, written in TypeScript on Bun. It mirrors ~/.claude additively — deletions upstream never propagate — indexes the result into SQLite FTS5, and serves a local explorer over it: wells, sessions read as blocks, token spend by day or model, the live agent roster. It is deterministic and model-free, and nothing leaves the machine.

Above the CLI sits an optional judgment layer: Claude sessions maintaining a private wiki from the mined material and promoting matured knowledge into git-committed docs — protocols, not code.

View source