Back to Projects
Agentic Cognitive Architecture

Agentic Temporal Memory & Episodic Graph Platform

Long-term cognitive memory framework modeling Ebbinghaus exponential forgetting curves, Neo4j temporal graph schemas, and Qdrant vector retrieval to slash prompt costs by 74%.

📖 Project Overview

Autonomous AI agents operating across multi-day enterprise workflows suffer from either catastrophic context memory loss (amnesia between execution sessions) or context window bloat (stuffing 50+ turns of verbose conversation logs, causing reasoning degradation and prohibitive token costs).

This platform deploys a biologically-inspired Dual-Tier Temporal Memory Engine. Short-term interactions enter working memory and are evaluated for salience. Core entities, decisions, and outcomes are extracted into a Neo4j Episodic Knowledge Graph paired with dense embeddings in Qdrant.

An asynchronous background decay worker applies mathematical Ebbinghaus forgetting curves (\(R = e^{-t/S}\)), allowing transient noise to fade over time while frequently reinforced knowledge nodes gain elevated retention strength. This architecture achieves 94.2% cross-session recall accuracy while cutting prompt token volume by 74%.

⚙️ Deployment Checklist & Runbook

1

Configure Mathematical Ebbinghaus Decay Engine

Calculate retention scores based on elapsed time and reinforcement frequencies:

python
import math
from datetime import datetime

def compute_retention(initial_strength: float, last_accessed_hrs: float, reinforcement_count: int) -> float:
    # Ebbinghaus exponential decay with reinforcement stabilization:
    stability_factor = 24.0 * (1.0 + 0.5 * reinforcement_count)
    decay_rate = last_accessed_hrs / stability_factor
    retention_score = initial_strength * math.exp(-decay_rate)
    return round(retention_score, 4)
2

Ingest Episodic Memory into Neo4j Graph

Establish Cypher query graph relations between agents, decisions, and external resources:

cypher
MERGE (a:Agent {id: $agent_id})
CREATE (m:EpisodicMemory {
    id: randomUUID(),
    summary: $summary,
    timestamp: datetime(),
    salience: $salience_score,
    retention: 1.0
})
CREATE (a)-[:RECORDED_EXPERIENCE]->(m)
WITH m
UNWIND $entities AS entity
MERGE (e:Entity {name: entity.name})
CREATE (m)-[:REFERENCES_ENTITY {confidence: entity.score}]->(e)
3

Deploy Hybrid Retrieval & Consolidation CronJob

Run periodic background compaction job pruning sub-threshold memories into summarized clusters:

yaml
apiVersion: batch/v1
kind: CronJob
metadata:
  name: memory-decay-compactor
  namespace: agent-memory
spec:
  schedule: "0 */4 * * *"  # Every 4 hours
  jobTemplate:
    spec:
      template:
        spec:
          containers:
          - name: compactor
            image: pradeeptalari14/agentic-memory-worker:latest
            env:
            - name: NEO4J_URI
              value: "bolt://neo4j-cluster.agent-memory.svc:7687"
            - name: RETENTION_THRESHOLD
              value: "0.25"
          restartPolicy: OnFailure

🔄 Configuration & Environment Keys

Environment Variable Description Suggested Value
MEMORY_DECAY_HALF_LIFE_HOURS Base half-life for unreinforced memory nodes. 24.0
MIN_SALIENCE_PERSIST_SCORE Cutoff score required to promote interaction to long-term graph. 0.70
GRAPH_VECTOR_ALPHA_WEIGHT Weighting between Neo4j graph proximity (1.0) and Qdrant cosine similarity (0.0). 0.65
MAX_ACTIVE_MEMORY_NODES Maximum consolidated memory items injected into prompt context. 8

📊 Architectural Workflow

graph TD
    Agent([Agent Execution Loop]) -->|New Observation| WorkingMemory[Working Context Buffer]
    WorkingMemory -->|Salience Evaluator| SalienceFilter{Score > 0.70?}
    SalienceFilter -->|No| TempBuffer[Ephemeral Redis Buffer (1 Hr TTL)]
    SalienceFilter -->|Yes| GraphIngest[Episodic Graph Ingestor]
    GraphIngest -->|Entity Extraction| Neo4j[(Neo4j Knowledge Graph)]
    GraphIngest -->|Vector Embeddings| Qdrant[(Qdrant Vector DB)]
    DecayCron[4-Hour Decay Worker] -->|Ebbinghaus Decay Formula| Neo4j
    DecayCron -->|Retire Stale Nodes| Archive[(Cold Parquet Archive)]
    Agent -->|Query Long-Term Context| HybridRetriever[Hybrid Graph-Vector Search]
    HybridRetriever -->|Cypher + Vector Cosine| ContextCompressor[Context Window Compressor]
    ContextCompressor -->|High-Salience Context (-74% Tokens)| Agent

🛠️ Useful CLI Reference

# Test live temporal memory decay and salience scoring: curl -X POST http://localhost:8000/api/ai/memory/ingest \ -H "Content-Type: application/json" \ -d '{"session_id": "session-prod-9", "interaction": "Upgraded cluster to Kubernetes 1.30 with Karpenter"}' # Query active episodic memory count in Neo4j: cypher-shell -u neo4j -p secret "MATCH (m:EpisodicMemory) RETURN count(m);"
📋 Code copied to clipboard!