@cognitive-engine/memory

Episodic and semantic memory with decay, search, and consolidation for cognitive-engine.

Install

npm install @cognitive-engine/memory

Two Memory Systems

Episodic Memory

Stores interaction episodes — what happened, when, with whom. Supports natural forgetting via exponential decay.

import { EpisodicMemory, EpisodeExtractor } from '@cognitive-engine/memory'

const memory = new EpisodicMemory(store, embeddingProvider)
const extractor = new EpisodeExtractor(llmProvider)

// Extract episode from interaction
const episode = await extractor.extract('user-123', message, response)

// Store with embedding for semantic search
await memory.storeEpisode(episode)

// Search by semantic similarity
const results = await memory.search({
  userId: 'user-123',
  query: 'team collaboration issues',
  limit: 5,
})

// Build context for response generation
const context = await memory.getContext('user-123', 'How is the project going?')

Semantic Memory

Knowledge graph of facts with confidence tracking and source attribution.

import { SemanticMemory, FactExtractor } from '@cognitive-engine/memory'

const semantic = new SemanticMemory(store, embeddingProvider)
const factExtractor = new FactExtractor(llmProvider)

// Extract facts from conversation
const facts = await factExtractor.extract(message, response)
// [{ content: 'User works on backend team', confidence: 0.9, source: 'stated' }]

// Store facts
for (const fact of facts) {
  await semantic.storeFact('user-123', fact)
}

// Recall relevant facts
const relevant = await semantic.recall('user-123', 'team structure')

Memory Decay

Episodes decay over time using exponential forgetting. Recent memories are vivid; old ones fade unless reinforced by recall.

// Decay is automatic during search
// Recalled memories get a salience boost (reinforcement)
// Memories below threshold are effectively forgotten

License

Apache-2.0