FoldBack

Context compression for LLM agents. Fold it down, fold it back.

FoldBack shrinks what your agent sends to the model — JSON tool outputs, logs, search results — using content-preserving transforms. No row is dropped, no value is discarded, and the provider prompt cache keeps hitting.

from foldback import compress

result = compress(messages, model="gpt-4o")
result.messages       # same format, fewer tokens — send these to the model
result.tokens_saved   # tokens saved
result.ratio          # 0.45 == 55% saved; 1.0 == nothing changed
result.transforms     # e.g. ["json:columnar"]

Why it exists

Most "context compression" tools made two expensive mistakes:

1. They compressed conversation history, dropping old messages — which busts the provider prompt cache on every call. On Anthropic that's a 90% discount thrown away.
2. They dropped data and hoped the model would ask for it back via a retrieval tool. If it doesn't realize data is missing, you get a confidently wrong answer with no error.

FoldBack refuses both:

• Passthrough is sacred. Everything before the last cache_control breakpoint is forwarded as the same objects — never copied, never re-serialized — so caches keep hitting.
• Only the live zone is touched (the latest user message / tool result).
• Token-gated. A transform is applied only when it actually reduces tokens. Otherwise the original is returned unchanged.

Guarantees, stated honestly

Two transform categories with different promises:

Content	Transform	Guarantee
JSON array of uniform objects	columnar (keys written once, rows as JSON arrays)	Reversible — exact round-trip, proven by property tests. restore_columnar() reconstructs the original.
Logs	strip ANSI, run-length-collapse consecutive identical lines to (xN)	Normalizing — no textual content lost; removes non-semantic bytes. Not byte-reversible.
Plain text	trim trailing whitespace, collapse blank-line runs	Normalizing — words and punctuation untouched.

The columnar transform only fires on uniform-schema arrays, so each row maps back to its keys unambiguously and "1" (string) never collides with 1 (number). Mixed-schema arrays are left untouched rather than compacted lossily.

from foldback import compress, restore_columnar
# round-trip proof
compressed = compress(messages).messages
# any columnar block is exactly restorable:
#   json.loads(restore_columnar(block)) == original_rows

Measured savings

Reproduce with python benchmarks/run.py --model gpt-4o (exact gpt-4o tokens):

Workload	Before	After	Saved
API response (100 rows)	2,803	1,421	49%
Build log (200 lines)	2,729	499	82%
Code search (50 hits)	1,892	1,159	39%

No marketing numbers — these come straight from the benchmark script.

Install

The PyPI package is foldback-ai; the import name is foldback.

pip install foldback-ai                 # zero dependencies
pip install "foldback-ai[exact]"        # + tiktoken for exact token counts
pip install "foldback-ai[anthropic]"    # + Anthropic SDK for the wrapper
pip install "foldback-ai[openai]"       # + OpenAI SDK for the wrapper

from foldback import compress           # import name stays `foldback`

Use it

Inline:

from foldback import compress, CompressConfig

result = compress(messages, model="claude-sonnet-4-5")
# or with options:
cfg = CompressConfig(model="gpt-4o", min_savings=0.2)  # require >=20% win
result = compress(messages, config=cfg)

Drop-in SDK wrappers (system prompt / tool defs stay frozen → cache-safe):

from anthropic import Anthropic
from foldback.integrations import with_anthropic

client = with_anthropic(Anthropic())
client.messages.create(model="claude-sonnet-4-5", messages=[...])  # auto-compressed

from openai import OpenAI
from foldback.integrations import with_openai

client = with_openai(OpenAI())
client.chat.completions.create(model="gpt-4o", messages=[...])     # auto-compressed

Develop

pip install -e ".[dev,exact]"
pytest                       # tests + coverage
ruff check foldback tests    # lint
mypy foldback                # strict type-check
python benchmarks/run.py     # savings table
python examples/demo.py

Deliberately NOT built

A network proxy, SSE streaming parser, Bedrock/Vertex signing, message scoring / relevance, a HuggingFace compression model, lossy row-dropping with retrieval. FoldBack is a library you call before your own SDK call — so it can never corrupt the wire.

Roadmap

• Diff / patch compaction
• CSV / Markdown-table input detection
• Rust core for the columnar path (only if profiling demands it)

Author

Built and maintained by Sudarshan Chaudhari (@SUDARSHANCHAUDHARI, sunny.sudarshan@gmail.com) — SudarshanTechLabs, Bangkok.

If FoldBack saves you tokens, a ⭐ on the repo is appreciated.

_{Developed with Claude Code (Opus 4.8).}

Acknowledgments

FoldBack's design was informed by studying chopratejas/headroom — a more ambitious context-compression project. FoldBack deliberately takes a narrower, library-only path (no proxy, lossless-first, prompt-cache-preserving) to avoid the cache-busting and lossy-retrieval pitfalls that project documented in its own realignment notes. Credit to its authors for mapping the problem space.

License

Apache 2.0 © Sudarshan Chaudhari / SudarshanTechLabs. See LICENSE and NOTICE.