Local-first 🧠memory for Hermes Agent that lives in your Obsidian vault and remembers project context. Nightly 😴 dream passes turn repeat corrections into confirmed preferences with measurable confidence. Adapters ship for Claude Code, Codex, and OpenClaw, with an MCP server for anything else.
TypeScript
/**
* Surfacing-replay bench suite (t_107a3249, lane C).
*
* Replays a recorded prompt sequence (a committed session fixture or a
* vault's own `session_turn` records, both via `replay-fixture.ts`)
* through the SHIPPED decision core over the disposable fixture vault,
* and scores every decision with the same
* {@link scoreProactiveRecall} denominations the proactive-recall suite
* uses: a silent miss is `know_to_ask`, a spoke-when-nothing-to-say is
* `false_fire`, and the two rates share one denominator so an
* always-inject gate can never hide.
*
* Scope, stated plainly: this is a FLOOR-CALIBRATION suite. It drives
* the shipped decision core directly, so the bound it measures is the
* graded admission key (`surfacing-key.ts`) - sweeping the shipped
* calibration against a permissive one, via the run's `recallFloor`, is
* what shows whether that bound discriminates. The gate's STRUCTURAL
* short-circuits (`duplicate`, `slash_command`, `shell_command`,
* `empty`) are exercised for recorded prompts: conversion carries each
* prompt's recorded previous prompt as `previous_prompt`, and the
* proactive-recall arm runs `evaluateSurfacingGate` ahead of the
* decision core, exactly as the transports do. See `failure-modes.ts`
* for what the scorer denominates.
*
* Rides the suite registry and the existing phase pipeline unchanged:
* conversion is a pure ReplayFixture -> BenchFixture whose questions are
* `proactive_recall`, so checkpoint, resume, per-turn `injected_tokens`
* and the abstain-costs-zero invariant all come from the one harness.
*/
import { parseBenchFixture } from "./fixture.ts";
import type { BenchFixture, BenchQuestion } from "./types.ts";
import { loadReplayFixture, type ReplayFixture } from "./replay-fixture.ts";
/**
* Convert a replay fixture into the canonical bench fixture the unchanged
* `runMemoryBench` consumes: every prompt becomes one `proactive_recall`
* question carrying its expectation, and the session's notes become the
* disposable vault the decisions retrieve against. No new category: the
* pipeline's existing proactive-recall arm drives and scores the run.
*
* Empty notes survive conversion, matching the loader's own stance: a
* recorded session with no ingested notes has nothing to materialize, so
* its fixture carries the questions alone and the run judges them
* against an empty vault. Each question carries the recorded previous
* prompt as `previous_prompt`, so the run's structural gate sees the
* same context the live hook saw.
*/
export function replayToBenchFixture(replay: ReplayFixture): BenchFixture {
const questions: BenchQuestion[] = replay.prompts.map((prompt) => ({
id: prompt.id,
category: "proactive_recall",
prompt: prompt.text,
expect_inject: prompt.expectInject,
...(prompt.previousText !== null ? { previous_prompt: prompt.previousText } : {}),
}));
// Validate through the canonical parser so a converted fixture is
// guaranteed to be exactly what runMemoryBench accepts. The empty-notes
// relaxation is named at the parser, not smuggled past it: a session
// with no ingested notes is questions-only by construction, and every
// replay entry point (`parseReplayFixture`, the vault-record
// conversion) may legitimately produce one.
return parseBenchFixture(
{
name: replay.name,
...(replay.description !== undefined ? { description: replay.description } : {}),
notes: replay.notes.map((note) => ({ path: note.path, body: note.body })),
continuity: [],
questions,
},
{ allowEmptyNotes: true },
);
}
/** The registry loader: a committed replay fixture file to a BenchFixture. */
export function loadReplayBenchFixture(path: string): BenchFixture {
return replayToBenchFixture(loadReplayFixture(path));
}