@timps-ai/memory-core

@timps-ai/memory-core

Persistent memory engine for AI agents — working, episodic, semantic, procedural memory with 17 intelligence tools, contradiction detection, bug pattern tracking, and burnout monitoring.

npm install @timps-ai/memory-core

Quickstart

import { MemoryEngine, InMemoryBackend } from '@timps-ai/memory-core'

const engine = new MemoryEngine('./my-project', {
  backend: new InMemoryBackend()
})

// Store a memory
engine.store({ content: 'The API uses JWT authentication', type: 'fact', tags: ['auth', 'security'] })

// Recall memories
const results = await engine.recall('how does authentication work')
console.log(results)

Architecture

22-layer memory architecture with 4 core memory types + 18 forge layers:

Layer	Name	Purpose
L1	Working Memory	Active context, goals, current task tracking
L2	Episodic Memory	Session history, interaction records
L3	Semantic Memory	Facts, preferences, conventions, patterns
L4	Procedural Memory	Reusable workflows, recipes
L5	ChronosForge	Causal graph tracking decisions and consequences
L6	ResonanceForge	Harmonic oscillator pattern resonance detection
L7	EchoForge	Reservoir computing with BFS context propagation
L8	AetherForgeERL	Epistemic Resonance Lattice
L9	HarmonicSheafWeaver	Sheaf-cohomology algebraic contradiction detection
L10	EngramLog	Immutable hash-chained audit trail
L11	ConsolidationEngine	Episodic → semantic memory promotion
L12	SynapticPruner	Active forgetting by importance scoring
L13	ProvenanceForge	Source tracking and chain of custody
L14	SpacedRepetitionForge	SM-2 scheduling for review timing
L15	ConstitutionalGuard	Prevents low-confidence writes
L16	AuditForge	Memory health reports and drift detection
L17	ProspectiveTrigger	Conditional recall ("when X happens, surface Y")
L18	BiasRevealer	Over/under-representation analysis
L19	ContextVector	State-dependent recall encoding
L20	RehearsalEngine	Spaced retrieval practice scheduling
L21	SchemaDistorter	Bartlett schema-driven distortion detection
L22	ConfidenceCalibrator	Multi-signal confidence scoring

Additional forge subsystems (SupraSheaf, QPTW, TitanicForge, QERW, QISRD, EclipseForge, QITRL) are available for advanced use cases.

Intelligence Tools (25 total)

• ContradictionDetector — detect conflicting stored facts
• BurnoutSeismograph — monitor developer fatigue signals
• RegretOracle — identify past decisions that caused issues
• TechDebtSeismograph — forecast maintenance burden
• BugPatternProphet — predict bug-prone patterns
• APIArchaeologist — track API quirks and gotchas
• VelocityTracker — monitor workflow velocity and patterns
• ArchitectureDriftDetector — track design divergence
• PatternLearner — extract reusable knowledge
• MeetingGhost — extract commitments from conversations
• DeadReckoning — simulate past decisions before repeating them
• LivingManifesto — derive values from behavioral patterns
• RelationshipIntelligence — track relationship health
• SkillShadow — identify tacit skills and gaps
• CurriculumArchitect — build personalized learning paths
• CodebaseAnthropologist — analyze codebase culture and norms
• InstitutionalMemory — capture departed contributor knowledge
• FalseMemoryDetector — detect hallucinated or incorrect memories
• ConfidenceCalibratorTool — score memory reliability
• SourceAttributor — trace memory origins
• ConflictResolver — resolve conflicting entries
• MemoryAuditor — full memory system audit
• ProspectiveTriggerTool — set conditional recall triggers
• BiasRevealerTool — surface systematic biases
• SchemaInferrer — derive structure from data

Storage Backends

Backend	File	Description
FileBackend	backends/FileBackend.ts	Local filesystem (default)
InMemoryBackend	backends/InMemoryBackend.ts	In-memory (testing)
PostgresBackend	backends/PostgresBackend.ts	PostgreSQL with RLS
SQLiteBackend	backends/SQLiteBackend.ts	SQLite (WAL mode)
RedisBackend	backends/RedisBackend.ts	Redis key/value
QdrantBackend	backends/QdrantBackend.ts	Vector search

Note on Intelligence Tool Results

New engines with no accumulated data will return empty/undefined fields from tools like burnout analysis, bug pattern prophecy, and architecture drift detection. These tools require multiple sessions of stored episodes and patterns before producing meaningful output.