compact

Compact: Context Compression

Compaction removes irrelevant lines from chat history and code context at 33,000 tok/s. Every line that survives is byte-for-byte identical to the input. Typical reduction is 50-70%.

Why not just truncate

Truncation drops the oldest tokens. The problem is that "oldest" and "least relevant" are different things — the file path the agent found on turn 3 is exactly what it needs on turn 90, and a sliding window throws it away first.

Compaction scores every line against a query and drops the ones the query doesn't need, wherever they sit in the history. Feed the same transcript in twice with two different queries and you get two different outputs — see query-conditioned for a runnable demo of that.

It is also not summarization. Summarization rewrites, which means drift and the occasional invented detail. Compaction only ever deletes whole lines, so an identifier that survives is still spelled the way your code spells it.

The one input this is wrong for: minified code, single-line JSON blobs, anything where most of the payload lives on one line. There is nothing for the model to cut. Split it into lines first.

Quickstart

curl -X POST https://api.morphllm.com/v1/compact \
  -H "Authorization: Bearer $MORPH_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": "def hello():\n    return 1\n\ndef unused():\n    pass\n\ndef world():\n    return 2",
    "query": "hello function",
    "compression_ratio": 0.5,
    "preserve_recent": 0
  }'

Two surfaces

Surface	Shape	Use when
POST /v1/compact	Native Morph. Returns output, per-message compacted_line_ranges, and usage.compression_ratio	You want the line ranges and the achieved ratio back
POST /v1/chat/completions with model: "morph-compactor"	OpenAI Chat Completions	You already have an OpenAI client and want a one-line change

POST /v1/responses also accepts model: "morph-compactor" for the OpenAI Responses API shape.

Parameters

Parameter	Type	Default	Description
input	string or array	—	Text or {role, content} array. One of input/messages required
messages	array	—	{role, content} messages. Takes priority over input
query	string	auto-detected	What matters for the next LLM call. Always set it
compression_ratio	float	0.5	Target fraction to keep. 0.3 aggressive, 0.7 light
preserve_recent	int	2	Keep the last N messages uncompressed
compress_system_messages	bool	false	Compress system messages too
include_line_ranges	bool	true	Return compacted_line_ranges
include_markers	bool	true	Emit (filtered N lines) markers where lines were cut

Wrap anything that must never be dropped in <keepContext> / </keepContext> tags, each on its own line. Preserved regions come back in kept_line_ranges.

Examples

Example	Language	Description
basic	TypeScript, Python	Compact a transcript, print before/after token counts
query-conditioned	TypeScript	One transcript, two queries, two different outputs
chat-completions	TypeScript	The morph-compactor model through the OpenAI SDK
websearch	Python (notebook)	Compact fetched web pages, then judge answer quality with and without
integration-prompt	—	Copy-paste prompts that make a coding agent wire compaction into your codebase

Practical notes

• Compact before the LLM call. Compacting a response after generation saves storage, not inference cost.
• Set preserve_recent to at least 3 in production. The examples here use 0 so the whole input is eligible.
• compression_ratio is a target, not a cap. When more of the input is relevant to the query, less gets cut.