Skip to content
MemberJunction

@memberjunction/context-crush

Dependency-light, framework-agnostic token-optimization primitives for AI agents and any TypeScript code that ships large payloads into a model’s context window. Every primitive is a pure function over plain strings/objects — no MemberJunction entity coupling, no AI calls, deterministic output.

What’s inside

Section titled “What’s inside”
ExportPurpose
CrushJSON(value, opts?)Structural JSON compression — collapses arrays-of-objects to a columns/rows table, elides null/empty fields, interns repeated long strings, and emits byte-stable minified output.
DescribeCrush(result)Renders a compact, model-readable legend explaining a CrushJSON result so the meaning is recoverable.
PartitionStablePrefix(messages, opts?)Splits a conversation into a byte-stable stable prefix and a mutable volatile tail, so callers can prune/compact without invalidating a provider’s KV-cache prefix.
CrushCode(source, lang, opts?) (subpath @memberjunction/context-crush/code)AST-aware source reduction — keeps signatures/structure, collapses non-focal function bodies. Behind a subpath export so the base import stays parser-free.

Usage

Section titled “Usage”
import { CrushJSON, DescribeCrush } from '@memberjunction/context-crush';


const result = CrushJSON(largeArrayOfRecords, { MaxChars: 8000 });
const prompt = `${DescribeCrush(result)}\n${result.Text}`;
// result.OriginalChars / result.CrushedChars report the savings.
// Code reduction lives behind a subpath so the base package stays zero-dep.
import { CrushCode } from '@memberjunction/context-crush/code';

Determinism

Section titled “Determinism”

CrushJSON produces byte-identical output for the same logical input (stable key ordering, no Date.now/random). This is load-bearing: it lets PartitionStablePrefix keep a re-rendered payload from breaking provider KV-cache hits across turns.

Credits & prior art

Section titled “Credits & prior art”

Several primitives here are independent TypeScript re-implementations of concepts from Headroom (chopratejas/headroom), an Apache-2.0 context-compression system for AI agents:

  • CrushJSON / DescribeCrush ← Headroom’s SmartCrusher
  • PartitionStablePrefix ← Headroom’s CacheAligner
  • CrushCode ← Headroom’s CodeCompressor

These are clean-room implementations written from the published algorithm descriptions — no Headroom source is copied, vendored, or depended upon, and this package does not use the “Headroom” name in its own identity. See the NOTICE file and the canonical attribution record in plans/agent-token-optimization.md (§0).