lossless-core

The engine behind pi-lossless: Lossless Context Management for a long-running agent. It stores every message verbatim, replaces aged context with searchable pointers, and keeps the original text readable by id, entry, or search. The algorithm is Lossless Context Management, by Clint Ehrlich and Theodore Blackman (Voltropy PBC).

It names no agent SDK. A host implements two ports, and the engine does the rest.

import { LcmStore, runCompaction, runRetrieval } from "lossless-core";
import type { ModelHost, Reader } from "lossless-core";

What it does

• Ingest. Every message lands in SQLite with its entry id, role, timestamp, and any file handle the host attached.
• Compaction. When occupancy crosses a swap ratio, a pass summarizes spans into leaf and condensed nodes and leaves a pointer in their place. Nothing is deleted: the store still holds the verbatim text.
• Retrieval. A reader answers a question by calling lcm_grep, lcm_describe, and lcm_expand, so the answer cites what was read.
• Redaction. Secret-shaped text is masked on its way to a model and never at the door, so an export or a read can still return what was stored.

The two ports

ModelHost is what the engine calls a model through: session(), find(), complete(), plus where notices go. Reader is one retrieval conversation: begin() returns a Reading whose next() appends results and returns the model's reply.

| Type | Question it answers | | ----------- | ----------------------------------------------------- | | ModelHost | Which model, and what did it answer | | Reader | What did the reader ask for, and what did it conclude |

src/host.ts documents both. The engine's boundary test fails any module under src/ that names an agent SDK, a bare specifier other than node:* and typebox, or a file outside the engine.

Where the rest lives

• pi-lossless: the Pi adapter, and every user-facing command.
• The docs site: how LCM behaves in a session, the config keys, and the limits.