Memory System

Five levels of intelligent memory for AI agents. Enable learning, context awareness, and organizational knowledge sharing across your platform.

Input
Process
Analyze
Output
Deep Learning Architecture

Features

Five Levels of Memory

From working memory to organizational knowledge

Working Memory

Per-run scratchpad. Cleared after execution completes. In-memory only for speed.

Short-term Memory

TTL-based cache persisting across runs. Auto-expires after configured duration.

Episodic Memory

Automatic run summaries with importance scoring. Tracks outcomes, decisions, and recall count.

Semantic Memory

Permanent knowledge: procedures, facts, policies. Versioned with full history.

Organizational Lessons

LLM-extracted lessons from successful runs. Hierarchical matching: agent → template → product → org-wide.

Automatic Injection

Relevant lessons injected into system prompts before execution. Use-count tracking for ranking.

Advanced Capabilities

Intelligent memory management and retrieval

Lesson Hierarchy

Lessons match by specificity: agent-specific → template-shared → product-shared → org-wide.

Context-Aware Scoring

Lessons scored by workflow DAG signature, integration context, and use-count. Most relevant first.

Auto-Provisioning

Default namespaces created per org. Default permissions granted per agent on creation.

Goal Memory Integration

Goal context injected into working memory. Goal episodes recorded. Success patterns extracted.

Memory Flow During Execution

How lessons are created, stored, and reused

Before Run

  • 1.Query organizational lessons matching context
  • 2.Load semantic memory (procedures, facts)
  • 3.Check short-term cache (avoid API calls)
  • 4.Inject lessons into system prompt

During Run

  • 5.Initialize working memory (scratchpad)
  • 6.TDAO loop reads/writes working memory
  • 7.Update short-term cache as needed
  • 8.Reference semantic memory for procedures

After Run

  • 9.Cleanup working memory (deleted)
  • 10.Record episodic memory (run summary)
  • 11.LLM extracts organizational lessons
  • 12.Score lessons by confidence & category

Example: Lesson Extraction & Reuse

Run #1: Review Vendor Contract

Discovered: "Check for auto-renewal clauses in Section 8"

Category: legal | Confidence: 0.92

Run #2: Review SaaS Agreement

Lesson injected from Run #1

Result: Flagged hidden 3-year auto-renewal clause

Ready to Empower Your Agents?

Build AI agents that learn and improve over time.