Topic-Aware Pruner

Detects topic boundaries in user input and emits memory.topic_shift_detected. The pruner does not itself rewrite history; it surfaces the shift so other plugins (summary buffer, compaction) can react.

Details


ID	`nexus.memory.topic_pruner`
Capabilities	none
Dependencies	none; uses `embeddings.provider` opportunistically when one is registered.

Two signals are combined:

Explicit phrase — substring match against a configurable list ("different question", "new topic", "let's move on", etc.). Cheap, deterministic. Lead-anchored phrases ("unrelated:", "separately,") are tagged user_explicit; substring matches are tagged phrase.
Embedding similarity — cosine similarity between the latest user input and the rolling centroid of the current topic’s user inputs. Below the configured threshold flags a shift. Runs only when an embeddings.provider is active. Tagged embedding.

Configuration

Key	Type	Default	Description
`enabled`	bool	`true`	Toggle the pruner.
`similarity_threshold`	float	`0.55`	Cosine similarity below which a new user input flags a topic shift. Used only when an `embeddings.provider` is active.
`keep_last_topic_full`	bool	`true`	Reserved for downstream consumers.
`explicit_phrases`	[]string	see code	Lowercase substrings that signal a topic shift. Replacing the list disables the defaults.

Defaults for explicit_phrases:

"different question", "different topic", "new topic", "new question",
"let's move on", "moving on", "change of subject", "switching gears",
"unrelated:", "separately,", "on a different note"

Events

Subscribes To

Event	Priority	Purpose
`io.input`	60	Run the classifier on every user input
`agent.turn.end`	60	Track the current turn for debouncing

Emits

Event	When
`memory.topic_shift_detected`	Per detected shift (carries `from_turn`, `to_turn`, `similarity`, `signal`)
`memory.curated`	Envelope event (`Layer: "topic_pruner"`, `CacheInvalidates: false`)
`embeddings.request`	Pointer-payload request when an `embeddings.provider` is active

Same-turn duplicate signals are debounced — at most one shift event per turn boundary.

Replay Determinism

Topic-shift decisions are non-deterministic (heuristic + embeddings). Each decision is journalled as a memory.topic_shift_detected event (Idea 01) so replay reproduces the same boundaries.

Example Configuration

plugins:
  active:
    - nexus.agent.react
    - nexus.memory.topic_pruner
    - nexus.embeddings.openai   # optional; enables the embedding signal

  nexus.memory.topic_pruner:
    enabled: true
    similarity_threshold: 0.55
    explicit_phrases:
      - "different question"
      - "new topic"
      - "moving on"

Keyboard shortcuts

Nexus Documentation