API Reference

Cognitive architecture primitives built on flow-state-dev. Provides attention, memory, and identity domains for agentic workflows.

Import: Use subpath exports. @thought-fabric/core/attention, @thought-fabric/core/memory, @thought-fabric/core/identity.

---

attention

Relevance and salience for what the agent attends to.

filterRelevance(config)

Handler block factory. Deterministic keyword-based relevance filtering. No LLM. Fast. Removes or annotates items below a threshold using keyword overlap heuristics.

import { filterRelevance } from "@thought-fabric/core/attention";

const block = filterRelevance({ name: "filter", criteria: { ... } });

Returns a BlockDefinition (handler). Use as a step in a sequencer or as a tool.

scoreSalience(config)

Generator block factory. LLM-based salience scoring along configurable dimensions (goal relevance, recency, novelty, emotional weight).

import { scoreSalience } from "@thought-fabric/core/attention";

const block = scoreSalience({
  name: "salience",
  dimensions: { ... },
  weights: { ... },
  model: "gpt-5-mini",
});

Returns a BlockDefinition (generator). Output schema: scores, composite, ranking, itemScores.

---

memory

Working memory: observe, remember, tick, snapshot.

Blocks

Function	Kind	Purpose
workingMemoryCapture(config?)	sequencer	Bundled: observe → remember → tick. Primary entry point.
workingMemoryObserve(config?)	generator	LLM extraction. Output: observations array.
workingMemoryRemember(config?)	handler	Persist observations into the resource.
workingMemoryTick(config?)	handler	Advance decay clock, recompute salience.
workingMemorySnapshot()	handler	Read current entries and turn counter.
workingMemoryAdd(config?)	handler	Manual entry. No LLM extraction.

Resource

• workingMemoryResource — defineResource() for working memory. Use in flow sessionResources.
• workingMemoryResources — Pre-keyed { workingMemory: workingMemoryResource }.

Context

• workingMemoryContextFormatter — Context slot for generators. Assign to context: [workingMemoryContextFormatter].

Helpers (verb-first naming)

Function	Purpose
add	Add entry with auto-eviction at capacity
evict	Remove entry by ID
pin / unpin	Toggle pinned status
refresh	Update lastAccessedAtTurn
advance	Tick decay, recompute salience for all entries
items	Read entries sorted by salience
formatForContext	Format entries for LLM context

Pure math (no side effects)

• computeDecay(elapsed, strategy, rate) — Decay factor. Strategies: power-law, exponential, none.
• computeSalience(entry, currentTurn, decay) — importance × decay(elapsed).

---

identity (placeholders)

Wave 2 placeholders. Not yet implemented.

• perspective(config) — Perspective block.
• constitution(config) — Constitution block.

---

Usage

import { workingMemoryCapture } from "@thought-fabric/core/memory";
import { filterRelevance, scoreSalience } from "@thought-fabric/core/attention";

const pipeline = sequencer({ name: "pipeline", inputSchema: chatInput })
  .work((input) => input.message, workingMemoryCapture({ model: "gpt-5-mini" }))
  .then(chat);

const filter = filterRelevance({ name: "filter" });
const salience = scoreSalience({ name: "rank" });

See Memory for a full guide.