Compact context, still recall: how loomcycle saves 80 percent of your tokens

Loomcycle's compact-plus-recall pipeline reduces token usage by up to 80 percent in long-running agent conversations, according to the company's technical documentation. The system, which uses recap (L1) and stateful (L2) retention modes alongside a Recall tool, cuts a 200-turn chat to about 0.20 times the original tokens on task-shaped runs, while preserving recall of evicted spans through an embedded index and optional durable memory consolidation.

A long-running agent conversation costs tokens quadratically: every turn re-reads the whole prior transcript, so a 200-turn conversation reads its own history 200 times. Loomcycle's answer is a compact-plus-recall pipeline RFC CR plus RFC CT that keeps a run under a bounded token budget while a Recall tool retrieves any span the distillation dropped. Three retention modes stack: append is the baseline every turn kept, cost O T squared ; recap L1 preserves the reasoning trace as prose and evicts the rest of the transcript at each boundary; stateful L2 carries a JSON-schema-validated state Sigma forward and evicts everything else. A tier-routed context.mode auto picks the mode from the provider that resolved: schema-free recap on a local backend, structured stateful on a frontier API. Providers gained a Local capability to make that a routing fact rather than a guess. A stateful agent can propose the shape its own state should take through emit state propose schema; the proposal is inert until an operator adopts it. Every eviction hands its span to two consumers: a run-scoped embedded index the Recall tool queries during the same run RFC CT P1; free-text query, silently falls back to the agent's durable memory, returns originals verbatim , and opt-in the durable memory consolidator that turns the span into a fact for future runs RFC CT P2 harvest to memory; whole-batched extraction, not per-span, because RFC CU Probe 2 measured 0.75 vs 0.00 on coreference-dependent facts . A stateful sub-agent hands its Sigma up to the parent as structured JSON on both single spawn and parallel spawn, so a fan-out that used to be N growing transcripts becomes N compact structured results. Recall auto-grants at every toolset-resolution site RunOnce, HTTP run path, session-continue, sub-agent spawn, resume so an operator does not have to remember to add it to the allowlist. Working range on a 200-turn chat: about 0.30 times the tokens at recap; about 0.20 times at stateful on a task-shaped run, which is the 80 percent ceiling. State diagrams for the retention lifecycle, the recall pipeline, and the harvest bridge.