openmemo-openclaw 2.8.2

OpenMemo Cognitive Memory Adapter for OpenClaw Agents

openmemo-openclaw

Memory Infrastructure for OpenClaw AI Agents

---

🎮 ▶ Try the Live Interactive Demo

See what happens to your AI agent after pip install openmemo-openclaw

Install → First Task → Workflow Reuse → Optimization — animated, step by step

Zero setup · Runs in your browser

---

What is openmemo-openclaw?

openmemo-openclaw is the official adapter that connects OpenMemo (the open-source Memory OS) to OpenClaw-based AI agents.

After one install, your agents:

• Remember past tasks, decisions, preferences, and workflows
• Avoid repeating completed work via task deduplication
• Automatically reuse successful playbooks
• Show memory cards in real-time as memory is created, reused, or optimized
• Stream events across the full agent lifecycle (start → recall → store → complete)

---

Quick Start

pip install openmemo-openclaw

from openmemo_openclaw import OpenMemoAdapter

# Initialize — memory starts automatically
adapter = OpenMemoAdapter()

# ✅ Write memories (tracked by Inspector, Policy Engine filtered)
adapter.write_memory("User prefers concise responses", scene="chat", memory_type="preference")
adapter.write_memory("Deployment uses Docker Compose", scene="devops", memory_type="workflow")

# ✅ Recall context for your agent
context = adapter.get_context("What are the deployment steps?")

# ✅ Full lifecycle hooks (auto-write + auto-recall)
adapter.on_user_message("Deploy my app")
adapter.on_task_complete("Deployed successfully via Docker Compose")

That's it. On startup, the adapter automatically:

1. Detects any memories written before it was installed → silently imports them
2. Intercepts all memory writes → ensures every write is tracked and visible in Inspector
3. Opens the Memory Inspector at http://localhost:8765/inspector

Important: Always use adapter.write_memory() to write memories — never call the OpenMemo SDK directly. Direct API calls bypass tracking and won't appear in the Inspector.

---

User Journey

After installing the adapter, agents go through a progressive memory experience:

Step 1: Install
  → Smart Memory Rules configured
  → Activation Panel shown to user

Step 2: First Task
  → Cold start (0 memory hits)
  → Task completes → memories stored
  → "Memory Created" card shown

Step 3: Next Similar Task
  → Warm start (memory recalled)
  → Workflow automatically matched
  → "Workflow Reused" indicator shown

Step 4: Optimization Detected
  → Repeated steps identified and skipped
  → "Optimization Detected" card shown
  → Tokens saved, time reduced

---

Architecture

┌─────────────────────────────────────────────────────────┐
│                    OpenClawAdapter                       │
│                                                          │
│  on_user_message() ──► PolicyEngine ──► _do_write()     │
│         │                                    │           │
│         ▼                                    ▼           │
│   EventBridge.emit()              ActivationStore.hook() │
│         │                                    │           │
│         ▼                                    ▼           │
│   UIController.show_*()           InsightCard generated  │
│                                                          │
│  Inspector API (port 8765)                               │
│    /events/recent  /ui/cards  /demo/simulate             │
└─────────────────────────────────────────────────────────┘

Core Layers

Layer	Module	Purpose
Adapter	adapter.py	Main entry point, wraps agent functions
Memory Policy	memory_policy/	Smart rules for what to store vs ignore
Activation System	activation/	Lifecycle hooks → insight cards
Event Bridge	event_bridge.py	Pub-sub event bus across all hooks
UI Controller	ui_controller.py	Renders memory cards in real-time
Inspector	inspector.py	HTTP server + demo simulation API

---

Memory Policy Layer

The Memory Policy layer decides what to store and what to ignore.

Default Smart Memory Mode

# What gets stored by default
STORE:
  - successful workflows
  - user preferences
  - key decisions
  - task patterns

IGNORE:
  - failed/noisy signals
  - temporary context
  - duplicate entries

Custom Rules

from openmemo_openclaw.memory_policy import PolicyEngine, RuleStore

rule_store = RuleStore()
rule_store.add_rule({
    "condition": "task_status == 'success' and memory_type == 'workflow'",
    "action": "store",
    "priority": 1,
})

engine = PolicyEngine(rule_store=rule_store)
adapter = OpenClawAdapter(policy_engine=engine)

Inspector API

GET  /policy/current        → current policy config
POST /policy/generate       → auto-generate rules from history
POST /policy/update         → update active rules

---

Activation System

The Activation System fires lifecycle hooks and generates UI insight cards at key moments.

Hooks

Hook	Trigger	Card Generated
on_install()	Adapter initialized	Activation Panel + Rule Card
on_task_start()	First-ever task	Guided First Task card
on_task_complete()	Task finishes with memories	Memory Created / Optimization Detected
on_workflow_adopted()	Past workflow reused	Reuse Indicator card

Example

from openmemo_openclaw.activation import ActivationStore

activation = ActivationStore()

# First task
guidance = activation.on_task_start(task_id="t-001", query="Deploy backend")

# Task complete — insight card generated automatically
insight = activation.on_task_complete(
    summary="Backend deployed via Docker Compose",
    memories_stored=3,
    workflows_reused=0,
    task_id="t-001",
)
# insight = {"type": "memory_created", "title": "✨ Memory Created", ...}

Inspector API

GET /activation/status       → current activation state
GET /activation/first-task   → first-task guidance data

---

Event Bridge

The Event Bridge is a pub-sub event bus that emits structured events across the entire agent lifecycle.

Event Types

Event	When
on_task_start	Agent begins a task
on_memory_recall	Memory retrieval occurs
on_memory_store	New memory written
on_workflow_adopt	Past workflow matched + applied
on_task_complete	Task finishes
on_policy_block	Policy engine blocks a write

Usage

from openmemo_openclaw.event_bridge import EventBridge

bridge = EventBridge()

# Subscribe to events
@bridge.subscribe("on_task_complete")
def handle_complete(event):
    print(f"Task {event['task_id']} done: {event['metadata']}")

# Emit from your agent
bridge.emit("on_task_start", task_id="t-001", metadata={"query": "Deploy backend"})
bridge.emit("on_memory_recall", task_id="t-001", metadata={"count": 3})
bridge.emit("on_task_complete", task_id="t-001", metadata={"status": "success"})

Inspector API

GET /events/recent?n=100    → recent events list
GET /events/stats           → event counts by type

---

UI Controller

The UI Controller renders memory cards in real-time as agents work.

Card Types

Type	Color	Trigger
activation_panel	🟢 Green	System initialized
rule_card	🔵 Blue	Memory rules configured
guided_first_task	🟣 Purple	First task detected
insight_memory_created	🩵 Cyan	Memories written after task
insight_optimization_detected	🟡 Amber	Workflow reuse saved steps
reuse_indicator	🟡 Amber	Past workflow matched

Usage

from openmemo_openclaw.ui_controller import UIController
from openmemo_openclaw.activation import ActivationStore

ui = UIController()
activation = ActivationStore()

# After task completes
insight = activation.on_task_complete(
    summary="Task done",
    memories_stored=2,
    workflows_reused=1,
    steps_skipped=2,
)
card = ui.show_insight(insight)
# card = {"type": "insight_optimization_detected", "title": "🚀 Optimization Detected", ...}

# Get current card
current = ui.get_current_card()

# Get all cards
all_cards = ui.get_all_cards(limit=20)

Inspector API

GET /ui/current             → latest card
GET /ui/cards?n=30          → all cards gallery

---

Memory Inspector

The Inspector is a built-in monitoring dashboard running at http://localhost:8765.

Sections

Section	What it shows
Impact Metrics	Memories stored, recalled, reused, tokens saved
Memory Experience Demo	Run a full simulated 4-step agent flow
Event Stream	Live event feed with color-coded badges
UI Cards Gallery	All generated memory cards
Task Inspector	Task trace + memory timeline
Top Memories	Highest-quality stored memories

Run the Demo

Open the Inspector at /inspector and click ▶ Run Demo to see:

1. Step 1 — Install: Memory Rules configured, Activation Panel shown
2. Step 2 — First Task: Cold start, memory created, insight card emitted
3. Step 3 — Workflow Reuse: Warm start, matching workflow applied, reuse card
4. Step 4 — Optimization: Steps auto-skipped, efficiency gain calculated

Inspector API

POST /demo/simulate         → run full 4-step simulation
POST /demo/reset            → clear simulation state
GET  /events/recent?n=100   → live event stream
GET  /ui/cards?n=30         → UI cards gallery
GET  /policy/current        → memory policy config
GET  /activation/status     → activation state

Start the Inspector

from openmemo_openclaw import OpenClawAdapter

adapter = OpenClawAdapter()
adapter.start_inspector(port=8765)
# → http://localhost:8765

---

Complete Example

from openmemo_openclaw import OpenClawAdapter

# Initialize — memory rules auto-configured, activation panel shown
adapter = OpenClawAdapter(api_key="your-key")
adapter.start_inspector(port=8765)

def my_agent(message: str) -> str:
    # your agent logic here
    return f"Result for: {message}"

# First task — cold start, memory stored after
result = adapter.on_user_message(
    user_message="Deploy backend service using Docker Compose",
    agent_func=my_agent,
)
# → "Memory Created" card emitted

# Second similar task — warm start, workflow reused
result = adapter.on_user_message(
    user_message="Deploy backend service again",
    agent_func=my_agent,
)
# → "Workflow Reused" card emitted
# → "Optimization Detected" card emitted (steps skipped)

---

What's Different vs Chat History / RAG

Feature	Chat History	RAG	openmemo-openclaw
Persistent memory across sessions	❌	❌	✅
Task deduplication	❌	❌	✅
Scene-aware recall	❌	❌	✅
Smart write policy (store/ignore rules)	❌	❌	✅
Lifecycle UI cards	❌	❌	✅
Live event stream	❌	❌	✅
Memory Inspector dashboard	❌	❌	✅

---

Changelog

v2.8.2 — Inspector UI Polish (Remove Demo Section, Fix Version Display)

• Removed: "MEMORY EXPERIENCE DEMO" section from Inspector — it was a development artifact not relevant to production use
• Fixed: Inspector header now shows openmemo-openclaw version (e.g. v2.8.2) instead of the underlying openmemo core version (v0.23.1)
• Fixed: Empty state messages in Event Stream and UI Cards no longer reference "Run the demo"

---

v2.8.1 — Cross-Version Compatibility Fix (LibraryProvider Resilience)

Fixes adapter failure when used with openmemo 0.23.1 — the current stable PyPI release.

Root Cause: LibraryProvider.recall_context() passed mode="kv" which does not exist in openmemo 0.23.1, causing a TypeError. The return value was also assumed to always be a dict, but some versions return a list.

Fixes:

• LibraryProvider.recall_context(): Progressive fallback through 4 call signatures; handles both dict and list return shapes
• LibraryProvider.write_memory(): Progressive fallback for different parameter signatures across openmemo versions
• LibraryProvider.search_memory() / list_scenes(): Wrapped in try/except with graceful empty-return fallback
• adapter.__init__(): _setup_write_intercept() now wrapped in try/except — interceptor failure is logged as a warning instead of crashing the adapter

Behavior: The adapter now works correctly with openmemo>=0.23.1. All degradation is logged at DEBUG/WARNING level; the adapter never crashes silently.

---

v2.8.0 — Zero-Friction Memory Tracking (Full 3-Layer Protection)

Eliminates the root cause: users can no longer accidentally write memories that are invisible to the Inspector, regardless of how they call the SDK.

Level 1 — Silent Startup Auto-Sync (_auto_sync_on_startup)

• On every adapter initialization, silently checks for pre-existing untracked memories
• Automatically imports them into FeedbackStore — zero user action, zero configuration
• Users upgrading from v2.7 or earlier: Inspector shows correct counts immediately on restart

Level 2 — Write Path Interceptor (_setup_write_intercept)

• Wraps _client.write_memory() at the instance level on startup
• Any write — whether via adapter.write_memory(), on_task_complete(), or direct client.write_memory() — is automatically tracked
• Uses _in_do_write flag to prevent double-counting from the internal _do_write() pipeline
• "Wrong path" no longer exists: all paths now lead to tracking

Level 3 — Quickstart Redesign (README)

• write_memory() is now the first example users see — correct pattern from day one
• Clear warning box: never call the OpenMemo SDK directly
• Startup behavior documented inline

v2.7.0 — Storage Health Guard (3-Layer Systemic Fix)

Solves the root problem where the Inspector shows 0 memories while memories were being written directly via API (bypassing OpenMemo tracking).

Layer 1 — Install-time Health Check (adapter.py)

• on_install() now runs _check_storage_consistency() automatically
• Detects mismatch between backend storage count vs. FeedbackStore tracked count
• Emits on_storage_mismatch event to EventBridge + logs a clear warning with fix instructions

Layer 2 — Inspector Storage Check + One-Click Sync (inspector.py + inspector_html.py)

• New GET /inspector/storage-check — returns {external_count, tracked_count, untracked_count, mismatch, status}
• New POST /inspector/sync — one-click import of all backend memories into OpenMemo tracking; metrics update immediately
• Inspector UI shows an amber warning banner when mismatch is detected; green "consistent" banner when healthy
• Auto-checks every 60 seconds; EN/ZH i18n supported

Layer 3 — Public Write Path Protection (adapter.py)

• New adapter.write_memory(content, scene, memory_type, metadata) — the correct, safe external write path
• All writes through this method are guaranteed to: (1) pass Policy Engine filter, (2) be tracked by FeedbackStore, (3) emit EventBridge events
• Fully documented with a clear warning against direct API writes

Also fixed: GET /inspector/metrics now returns both memories_stored and total_memories (alias), fixing a UI display inconsistency.

Migration for existing users:

# If you have memories in storage but Inspector shows 0:
# Option A: One-click via Inspector UI → visit http://localhost:8765/inspector
# Option B: API call
import requests
requests.post("http://localhost:8765/inspector/sync")

# Option C: Use write_memory() for all future external writes
adapter.write_memory("User prefers concise responses", scene="chat", memory_type="preference")

v2.6.0 — Event Bridge + UI Controller

• event_bridge.py: Unified pub-sub event bus — 7 event types (on_task_start, on_memory_recall, on_memory_store, on_workflow_adopt, on_task_complete, on_policy_block), subscribe() / emit() / get_recent() / get_stats()
• ui_controller.py: Real-time UI card renderer — 6 card types indexed by type, get_current_card() / get_all_cards() / get_stats()
• adapter.py: All hooks emit to EventBridge; UIController renders all card types; full lifecycle instrumentation
• Inspector API: GET /events/recent, GET /events/stats, GET /ui/current, GET /ui/cards
• Inspector HTML: Live Event Stream table + UI Cards Gallery + ▶ Run Demo interactive simulation button

v2.5.0 — Memory Policy + Activation System

• memory_policy/: policy_engine.py, rule_store.py, rule_generator.py — Smart Memory Mode with default store/ignore rules; _do_write() gate in adapter
• activation/: activation_store.py — 4 lifecycle hooks (on_install, on_task_start, on_task_complete, on_workflow_adopted), auto-generates InsightCards
• Inspector API: GET /policy/current, POST /policy/generate, POST /policy/update, GET /activation/status, GET /activation/first-task
• Inspector HTML: Activation Status section with Memory Rule Card + System Status + Insight Card

v2.4.0 — Inspector V3 + Feedback Tracking

• inspector_html.py: Full V3 UI — Impact Metrics, Task Trace, Timeline, Memory Details, Top Memories
• inspector.py: V3 API endpoints — /inspector/metrics, /task-trace, /timeline, /memory-profile, /top-memories
• feedback_store.py: Lightweight in-memory event tracking with per-task traces and quality scores (quality_score = adoption×0.4 + success×0.4 + recency×0.2)
• adapter.py: FeedbackStore integration + UserMemory extraction/recall hooks

v2.3.0 and earlier

See PyPI release history.

---

Installation

# First-time install
pip install openmemo-openclaw

# Upgrade to latest
pip install openmemo-openclaw --upgrade

Requirements:

• Python 3.9+
• openmemo >= 0.23.0

---

Contributing

We welcome contributions! Open an issue or submit a pull request.

---

License

MIT — see LICENSE

---

Built on OpenMemo — The Memory OS for AI Agents.