thefixer3x/behavior-memory

Skill: Behavior Memory Integration

Purpose

Teach agents to record successful workflows and recall them in future sessions using the LanOnasis MaaS API — so Claude learns from real execution history instead of starting from scratch every time.

Local-first: Use cached embeddings for recall. API is fallback, not primary.

---

API Quick Reference

Base URL: https://api.lanonasis.com/api/v1

| Operation | Endpoint | Method | |-----------|----------|--------| | Record pattern | /memories | POST | | Recall patterns | /memories/search | POST | | List patterns | /memories?type=workflow | GET | | Update pattern | /memories/{id} | PUT | | Delete pattern | /memories/{id} | DELETE | | Analyze patterns | /intelligence/analyze-patterns | POST | | Find related | /intelligence/find-related | POST | | Extract insights | /intelligence/extract-insights | POST |

Authentication

Primary:  Authorization: Bearer <oauth2_pkce_token>
Fallback: X-API-Key: lano_*
Legacy:   Authorization: Bearer <jwt>

---

Session Flow

SESSION START
  └─ 1. Recall: semantic search for patterns matching current task
  └─ 2. Inject: enrich agent context — "User usually does X when Y"

SESSION EXECUTION
  └─ 3. Act: execute with pattern-informed behavior, track tool outcomes

SESSION END
  └─ 4. Record: store trigger → actions → outcome as workflow memory

---

Recording a Pattern

POST /memories
{
  "title": "Workflow: Fix auth bug in TypeScript API",
  "content": "{\"trigger\":\"fix auth bug\",\"actions\":[{\"tool\":\"Read\",\"outcome\":\"success\"},{\"tool\":\"Edit\",\"outcome\":\"success\"},{\"tool\":\"Bash\",\"outcome\":\"success\"}],\"final_outcome\":\"success\",\"duration_ms\":45000}",
  "type": "workflow",
  "tags": ["auth", "bugfix", "typescript-api"],
  "metadata": {
    "confidence": 0.85,
    "use_count": 1,
    "context": {
      "directory": "/home/user/onasis-gateway",
      "project_type": "typescript-api",
      "branch": "main"
    }
  }
}

Confidence Formula

function calculateConfidence(outcome) {
  let score = 0.5;                              // base
  if (outcome.userExplicitApproval) score += 0.3;
  if (outcome.noErrors)             score += 0.1;
  if (outcome.completedAllSteps)    score += 0.1;
  return Math.min(score, 1.0);
}

Only record patterns with confidence ≥ 0.5. Never record failed workflows.

---

Recalling Patterns

POST /memories/search
{
  "query": "fix authentication bug in typescript",
  "type": "workflow",
  "threshold": 0.7,
  "limit": 5
}

Client Configuration (always use offline-fallback for recall)

const client = new MemoryIntelligenceClient({
  processingMode: 'offline-fallback',
  enableCache: true,
  cacheTTL: 300000  // 5 minutes
});

---

MCP Tools

Analysis (Read-Only)

// Analyze patterns over time
memory_analyze_patterns({ user_id, time_range_days: 30, response_format: "markdown" })

// Find related memories by vector similarity
memory_find_related({ memory_id, user_id, limit: 10, similarity_threshold: 0.7 })

// Extract themes and insights
memory_extract_insights({ user_id, topic: "deployment", memory_type: "workflow", max_memories: 20 })

Behavior Write Tools (v2.0)

// Record a successful workflow pattern
behavior_record({
  user_id,
  trigger: "fix auth bug",
  context: { directory: "/apps/mcp-core", project_type: "typescript-api", branch: "main" },
  actions: [{ tool: "Read", outcome: "success" }, { tool: "Edit", outcome: "success" }],
  final_outcome: "success",
  confidence: 0.85
})

// Recall relevant patterns for current context
behavior_recall({
  user_id,
  context: { currentDirectory: "/apps/mcp-core", currentTask: "fix auth middleware", projectType: "typescript-api" },
  limit: 5,
  similarityThreshold: 0.7
})

// Get AI-powered next-action suggestions
behavior_suggest({ user_id, current_state: { task, context, history } })

---

Memory Types

| Type | Use For | |------|---------| | workflow | Multi-step action sequences ← primary for behavior patterns | | context | Session context (directory, project type) | | project | Project-specific patterns | | knowledge | Learned preferences and rules | | personal | User-specific behavior preferences |

---

Failure Modes & Recovery

| Failure | Recovery | |---------|----------| | API unreachable | Use processingMode: 'offline-fallback' — serve from local cache | | Cache miss + API down | Proceed without pattern context; log the miss | | Auth error (401) | Try X-API-Key header fallback; if both fail, skip pattern recording | | Confidence < 0.5 | Do not record; log reason | | Duplicate detected (similarity > 0.95) | Update use_count on existing record instead of creating new |

---

Hard Rules

• Never store PII, tokens, or secrets in patterns
• Never record failed sessions as successful patterns
• Never call API when cache is fresh and relevant
• Never create duplicates — check first with detectDuplicates
• Never overwrite a high-confidence pattern with a lower-confidence one