Every LLM session gets
more expensive as it runs.
That's arithmetic. Not a bug.

Turn N replays the whole transcript as input, so turn N costs proportional to N and the session costs O(N²). That is arithmetic from the pricing page, not a property of any SDK.

Burnless stops replaying it. Not by asking the LLM to be careful. By separating Cognitive Execution (Workers) from State Management (Capsules/Maestro) — the same design move TCP/IP made when it separated applications from the network. Same move, not the same scale: TCP/IP defines internet infrastructure; Burnless is a small MIT Python framework that wires caching, tier routing and capsules together in a particular way. Workers receive the current task only. The Brain reads ~80-character capsules instead of transcripts, and the system prefix stays byte-identical, so the provider's cache stays hot.

Measured, not promised: a real paid API run — $5.76 of actual spend against claude-opus-4-7, raw response.usage, no mocks — cut turn-10 cost by 90.3% vs no-cache replay, and a further ~30% vs an already-cached replay baseline. The gap grows with session length. Reproduce it with your own key: python bench/run.py --turns 10.

Two receipts from the author's own machines: 121 million tokens kept out of context since 2026-05-18, logged turn by turn in an append-only JSONL — and one order-of-magnitude week, 97% of weekly quota down to 1% on comparable work. The first is a measurement. The second is an anecdote, labeled as one. Both are why this project exists.

Standalone O(N²) vs Burnless O(N) cost curve

Calibrated from real Anthropic API runs. Reproduce: python bench/v2.py --simulate

Modeled curve — calibrated from the real claude-opus-4-7 run below
 Turns   Standalone   Burnless   Savings
   2        $0.80       $0.14    82.7%
   5        $2.06       $0.29    86.1%
  10        $4.34       $0.54    87.6%
  20        $9.59       $1.07    88.9%
  50       $30.72       $2.83    90.8%
Baseline = no-cache full replay. If your setup already caches, the honest comparison is the run below. Derivation and failure cases: MATH.md.
Real API run · claude-opus-4-7 · 10 turns · $5.76 actual spend · no mocks
 A — No cache, full replay   $4.66
 B — Cache + full replay     $0.65   −86.0% vs A
 C — Burnless capsules       $0.45   −90.3% vs A
B is the honest baseline — most modern setups already cache. Burnless still beat it by ~30% at this length, and the advantage grows with the session. Raw response.usage, reproducible with your own key.
MIT licensed · Self-hosted · No backend required · pip install burnless
Docs history, on the record: early versions of this page overclaimed — a TCP/IP identity claim, a universal "16×". Publicly retracted 2026-05-08, receipts in git log. Every number here since is measured, reproducible, or labeled as the anecdote or extrapolation it is. The full note →
Roadmap: the same capsule memory across engines. Claude Code integration shipped; Codex lifecycle hooks in beta on the author's machine; cross-engine restore is roadmap, not a shipped claim.

Beyond cost. The protocol layer.

A token is not an abstraction. It is compute, and compute is electricity. If LLM inference reaches the 1–5% of global electricity that public projections suggest for the next decade, O(N²) replay stops being a pricing quirk and becomes a trajectory.

Burnless is a provider-agnostic layer that sits between any user and any LLM. The model receives a capsule. The user sends a message. The orchestration layer speaks compact state. The contracts — worker envelope, capsule format, tier semantics, audit gates — are written down in PROTOCOL.md, independent of the reference implementation: a protocol in the LSP sense — a specified interface others can implement — not a claim of TCP/IP-scale foundations.

60–90 TWh/year

Back-of-envelope extrapolation (VISION.md) if this pattern reached 1% of global LLM inference. Denmark consumes 35 TWh/year. An extrapolation, not a measurement — check the arithmetic yourself.

Structurally unblockable by design

A semantic capsule is a dense summary of session state sent as ordinary text. The protocol needs no provider-specific feature to carry compact state.

Open, on purpose.

A layer like this will exist. The question is who defines it and in whose interest. Burnless's answer: MIT, documented, contracts public so anyone can implement them.

Read the founding vision →

Why the curve is quadratic.

Every turn in a standalone agent loop replays the full conversation as input. Cost on turn N is proportional to N, so total cost across N turns is Θ(N²). That is arithmetic from the pricing page, not a property of any SDK.

1. Capsules, not transcripts

Brain history holds ~80-char summaries of each turn, not the raw exchange. Full output stays on disk, read on demand.

2. Hot prefix cache

System prompt is byte-identical every turn with cache_control, so it bills at cache-read price — 10% of base input on Anthropic. Caches are per provider and per model; nothing is "shared across models".

3. Tiers are roles, not models

Any model as Brain. Any model as Worker. GPT-4o, Opus, Sonnet, Codex, Ollama — one-line config change.

4. Living glossary

Core terms, project terms and session deltas form a compact language between human intent, Maestro and Workers.

5. Privacy by architecture

Run encoder/decoder locally → cloud sees only capsules, not raw text. Run Maestro locally → cloud workers see disconnected task fragments with zero conversation context. Four levels, one config line.

We don't ask you to trust the table. Reproduce it: python bench/run.py --turns 8 with your own API key.

How it looks.

A chat-first CLI. Slash commands users already understand. Compact state under .burnless/. No hosted backend.

# in any project
$ pip install burnless
$ burnless setup
$ burnless
   # Burnless Chat — /commands, /maestro, /model, /workers, /native

$ burnless delegate "fix the failing tests"
   → d001 routed to silver/codex  (matched: test)
$ burnless delegate "summarize the logs"
   → d002 routed to bronze/haiku  (matched: summarize)
$ burnless run d002
   heartbeat: reading · idle 7s  (ephemeral)
   OK:d002 — typed report + capsule saved

Gold/silver/bronze are quality/cost bands, not vendors. The default setup can mix Claude, Codex and Ollama; users can replace any tier in .burnless/config.yaml.

Cache compaction is realtime ROI math: Burnless freezes immutable blocks and only creates a new super-capsule when future cache-read savings beat the write+compaction cost.

Editions.

The protocol is MIT and stays free. Pay for the things organizations need to trust it at scale: governance, compliance, and team infrastructure.

Burnless

Free · MIT

  • Open protocol, MIT licensed
  • Full CLI, runs locally
  • Brain + Worker + capsule history
  • Shared prefix cache, realtime capsule compaction
  • Typed execution/thought reports with evidence audit
  • Ephemeral heartbeat progress without scroll-history noise
  • Local burnkey semantics planned in the protocol
  • Provider-agnostic — any LLM as Brain or Worker
  • Reproducible benchmark
View on GitHub

Burnless Cloud/Enterprise

Soon · waitlist

  • Shared cache across machines and teammates
  • KMS/HSM key custody and retention policy
  • Audit logs, legal hold and destruction reports
  • SSO + team permissions
  • Dashboards, SIEM/webhooks and priority support
Join Cloud waitlist

Don't trust the table. python bench/run.py --turns 8 with your own API key.

Burnless Cloud — waitlist.

Hosted features for teams: shared cache, dashboards, KMS/HSM key custody, audit logs, SSO and retention policy. The protocol is free and open source — Cloud/Enterprise is for organizations that need governance.