Mnemosyne.Pipeline.Episode (mnemosyne v0.1.6)

Functional core for managing in-progress episodes.

An episode tracks a sequence of observation-action steps organized into trajectories. Trajectory boundaries are detected when subgoal embedding similarity drops below a threshold.

Summary

Types

step()

A single observation-action step within an episode

t()

trajectory()

A completed trajectory segment

Functions

append(episode, observation, action, opts)

Appends an observation-action step, inferring subgoal, reward, and state via LLM.

build_trajectory_from_steps(steps)

Builds a trajectory struct from a list of steps sharing the same trajectory_id.

close(episode)

Closes the episode, grouping steps into trajectory segments.

new(goal)

Creates a new open episode targeting the given goal.

score_pending_reward(episode, opts)

Scores the reward for the last step using a sentinel next-observation. Call before close.

Types

step()

@type step() :: %{
  index: non_neg_integer(),
  observation: String.t(),
  action: String.t(),
  subgoal: String.t(),
  state: String.t() | nil,
  reward: float() | nil,
  embedding: [float()] | nil,
  trajectory_id: String.t()
}

A single observation-action step within an episode

t()

@type t() :: %Mnemosyne.Pipeline.Episode{
  closed: boolean(),
  current_subgoal_embedding: [float()] | nil,
  current_trajectory_id: String.t(),
  goal: String.t(),
  id: String.t(),
  steps: [step()],
  trajectories: [trajectory()]
}

trajectory()

@type trajectory() :: %{id: String.t(), steps: [step()], subgoal: String.t()}

A completed trajectory segment

Functions

append(episode, observation, action, opts)

@spec append(t(), String.t(), String.t(), keyword()) ::
  {:ok, t(), Mnemosyne.Notifier.Trace.Episode.t()}
  | {:error, Mnemosyne.Errors.error()}

Appends an observation-action step, inferring subgoal, reward, and state via LLM.

build_trajectory_from_steps(steps)

@spec build_trajectory_from_steps([step()]) :: trajectory()

Builds a trajectory struct from a list of steps sharing the same trajectory_id.

close(episode)

@spec close(t()) :: {:ok, t()} | {:error, Mnemosyne.Errors.Invalid.EpisodeError.t()}

Closes the episode, grouping steps into trajectory segments.

new(goal)

@spec new(String.t()) :: t()

Creates a new open episode targeting the given goal.

score_pending_reward(episode, opts)

@spec score_pending_reward(
  t(),
  keyword()
) :: {:ok, t()} | {:error, Mnemosyne.Errors.Invalid.EpisodeError.t()}

Scores the reward for the last step using a sentinel next-observation. Call before close.