Skip to main content

​
Memory Priority Matrix

PriorityScopeContentsUse Case
1user_idUser preferences, learned behaviorsPersonalization
2agent_idManowar execution patterns, skill learningsMulti-run improvement
3run_idCurrent execution contextWithin-run coherence
Search cascades from user → agent → run, ensuring the most relevant context surfaces first.

​
The Safe Wipe Pattern

When context window usage exceeds the dynamic threshold, Manowar compresses state without information loss: 
┌─────────────────────────────────────────────────────────┐
│  1. Summarize current context (coordinator model)       │
│     └── 200 words max, preserving key facts             │
│                                                         │
│  2. Store summary in Mem0 with graph extraction         │
│     └── Entities and relations automatically extracted  │
│                                                         │
│  3. Return [CONTEXT REFRESHED] marker                   │
│     └── State reset while summary available via search  │
└─────────────────────────────────────────────────────────┘
The dynamic threshold scales with model capacity: 
Threshold = 55% + 0.5 × log₁₀(contextWindow / 1024)
  • 32k context → ~56.8%
  • 128k context → ~58.9%
  • 4M context (MiniMax M2.1) → ~72.3%

​
Semantic Retrieval

Context retrieval uses e5-mistral-7b-instruct embeddings with Mem0’s advanced features:
  • Reranking: Relevance optimization after initial retrieval
  • Keyword Search: Hybrid semantic + lexical matching
  • Metadata Filtering: Type-based access (step_reflection, checkpoint_summary, trigger)

​
Checkpoint System

Internal agents annotate execution with ephemeral checkpoints:
TypePurpose
observationWhat the agent perceived
decisionChoices made and reasoning
outputResults produced
insightMeta-learnings for future runs
errorFailure modes for pattern avoidance
Checkpoints persist to Mem0 at workflow completion for cross-run learning.

​
Ambient Intelligence / Proactive Systems

Manowar supports autonomous operation through natural language triggers—enabling workflows that execute without human initiation.

​
Trigger Architecture

Natural Language → Pattern Matching → LLM Fallback → Cron Expression
   "every day          [matched]                       "0 9 * * *"
    at 9am"
Common patterns are parsed instantly:
  • “every 5 minutes” → /5 * * * *
  • “every Monday at 3 PM” → 0 15 * * 1
  • “daily at midnight” → 0 0 * * *
Complex patterns fall back to LLM parsing with nvidia/nemotron-3.

​
Trigger Storage

Triggers persist in Mem0 graph memory: 
{
  type: "trigger",
  trigger_id: "trg-...",
  manowar_wallet: "0x...",
  cron_expression: "0 9 * * 1-5",
  enabled: true,
  trigger_data: JSON.stringify(fullDefinition),
}
This enables:
  • Cross-session persistence
  • Graph-based discovery
  • Workflow-trigger relationships

​
Lifecycle Hooks

Beyond triggers, Manowar defines lifecycle hooks for reactive behaviors:
HookFires When
pre-executionBefore workflow starts
post-stepAfter each step completes
on-errorAny step fails
on-completeWorkflow succeeds
on-context-cleanupSafe wipe triggered
on-restartWorkflow resumes after cleanup
Hook actions include: webhook, agent, memory, notify, log.

​
Run Observability

Every execution is tracked with:
  • LangSmith Integration: Full trace visibility, token metrics, reasoning_tokens for thinking models
  • TrackedRun State: pending → running → success/error/cancelled
  • Cron Stats: Per-trigger execution history and success rates