mnemo-codemode 0.5.21

Code-mode recall for Mnemo (v0.4.0 P0-3) — agents call recall as a sandboxed WIT function instead of a JSON tool, dropping per-turn token cost by ~95%.
Documentation

v0.4.0 (P0-3) — Code-mode recall.

Cloudflare's Code Mode MCP (announced 2026-04-24) showed that agents calling tools through generated host-side code instead of JSON tool envelopes can drop per-turn token cost by ~99.9%. The same shape applies to recall: instead of the LLM paying for tool_call(recall, {query: "..."}) plus tool_result([memory, memory, ...]) JSON each turn, the host hands the LLM a sandboxed wasm host whose imports it can call as plain functions.

This crate ships:

  1. The host-side data shapes [CodeModeRecall], [RecallBundle], [ResourceBudget].
  2. A pure host-side runner [run_code_mode_host] that accepts a pre-built guest "program" (a list of recall calls) and produces a [RecallBundle] — used by the token-budget tests + the mnemo recall --code-mode CLI in the binary crate.
  3. The WIT world definition under wit/mnemo-memory.wit describing the import/export surface a real wasm guest must implement.

The wasmtime+wasi-stripping path lives behind the wasm feature and is wired in a follow-up; the host-side contract is fully tested today and is what mnemo-cli's recall --code-mode currently dispatches to.