Chrollo

A retrieval layer for Pi's native session history. Search past conversations, read matched windows. No storage, no capture, no injection.

Why

Pi records every session to disk as .jsonl. Nothing in Pi lets you search across that history when you're back in a conversation. Chrollo fills that one gap: find what you worked on before, when you ask for it.

It does not capture, summarize, store, or inject anything. Pi is the source of truth; Chrollo only reads it.

Quick Start

Requires ripgrep on your PATH.

sudo apt install ripgrep      # Linux
brew install ripgrep          # macOS

pi install npm:@k3_2o/pi-chrollo

Then in any Pi session, ask about something from a past conversation — the agent searches your history and reads back the relevant turns.

How It Works

Two tools, one workflow:

1. search_memory — search past sessions by keyword. Returns markers like path:line | role: preview, most-recent sessions first.
2. read_memory — read a bounded window around a marker, rendered readably ([HH:MM] role: text).

The agent calls search_memory when you reference past work, picks a marker, then calls read_memory with that marker's line number to see the surrounding context.

you:    "what did we decide about the k3s ingress?"
agent:  search_memory("k3s ingress")  → 15 markers
        read_memory(path, offset=77, limit=10)  → readable window

What It Does and Doesn't Do

| | | |---|---| | Searches | All of Pi's session history, by keyword | | Stores | Nothing. Reads Pi's existing .jsonl files | | Captures | Nothing. Pi already captures every turn | | Injects | Nothing. You decide when to recall | | Calls an LLM | Never | | Needs a daemon | No. One binary (ripgrep) + native Node |

Tool outputs and internal reasoning are filtered out automatically — only real conversation turns appear in results and reads.

Ranking

Chrollo keeps it boring on purpose.

• One ripgrep call. rg --json --sortr modified -m 200 -F -e <term> -i searches the corpus and orders sessions by recency (file mtime) in a single pass.
• Rank by coverage, not just recency. A line mentioning more of your distinct keywords ranks above a more recent line matching only one — so a distinctive old conversation surfaces over recent topical chit-chat. Ties fall back to recency order. No BM25, no stemming, no typo fallback, no global stats scan.
• Filter, then cap. Each matched line is structurally filtered (tool outputs, thinking, metadata dropped); the -m 200 per-file cap keeps a mid-file answer reachable before ranking, then a per-file diversity cap + MAX_RESULTS=15.
• Honest failure. A real timeout says "timed out — retry", never a fake "no memories". Esc genuinely cancels.

Working limit: an undistinguished keyword (one that also appears across many recent sessions) can't guarantee the "right" old session — that's inherent to keyword retrieval. The overlay pays off when the query has at least one term distinct to the target.

License

MIT — see LICENSE.