Openclaw Local Memory

🧠 Local Memory Plugin v0.4

A brain-like memory system for OpenClaw. Remembers what matters, forgets what doesn't, and builds a persistent understanding of you over time.

> Zero-config, no external service, no API key, works out of the box.

Features

🧠 Brain-Like Memory Architecture

🔍 Smart Recall

💾 Intelligent Capture

🗑️ Self-Maintaining

Tools

| Tool | Description | |------|-------------| | local_memory_search | Search memories by natural language (semantic) | | local_memory_store | Manually save a specific memory | | local_memory_list | List all memories, optionally filtered by category | | local_memory_profile | View user profile (entities, preferences, facts) | | local_memory_stats | View memory statistics | | local_memory_recent | Get recently accessed memories | | local_memory_forget | Delete memory matching a query | | local_memory_wipe | Delete ALL memories (irreversible) |

How It Works

Memory Lifecycle

1. Capture → User + Assistant exchange 2. Significance Assessment → Score based on patterns (decisions score high, greetings low) 3. Storage → If significant enough, store with extracted entities and tags 4. Importance Calculation → Based on category, length, entities, source 5. Decay Over Time → Importance decreases exponentially 6. Recall → On query, combine TF-IDF relevance + importance + recency 7. Pruning → When max reached, lowest combined-score memories removed

Recall Scoring Formula

score = (relevanceWeight × tfidf_similarity) 
      + (importanceWeight × decayed_importance)
      + (recencyWeight × recency_factor)

Significance Detection Patterns

| Pattern | Category | Weight | |---------|----------|--------| | entschieden, geplant, wird, werden | decision | 0.30 | | ich bin, mein, unser Unternehmen | identity | 0.25 | | bevorzug, immer, nie, prefer | preference | 0.25 | | api_key, password, token | credential | 0.20 | | skill, können, fähig | skill | 0.20 | | projekt, build, deploy | project | 0.15 |

Configuration

{
  "autoRecall": true,
  "autoCapture": true,
  "captureInterval": 8,
  "captureSignificantOnly": true,
  "minSignificanceScore": 0.5,
  "profileFrequency": 15,
  "includeProfileOnFirstTurn": true,
  "maxRecallResults": 5,
  "similarityThreshold": 0.35,
  "maxMemoryInjections": 3,
  "contextBudget": 2000,
  "maxMemories": 500,
  "pruneOlderThanDays": 30,
  "decayRate": 0.05,
  "chunkSize": 800,
  "importanceWeight": 0.25,
  "recencyWeight": 0.25,
  "relevanceWeight": 0.5
}

| Option | Default | Description | |--------|---------|-------------| | autoRecall | true | Inject relevant memories before each turn | | autoCapture | true | Auto-capture conversation exchanges | | captureInterval | 8 | Capture every N turns (higher = less storage) | | captureSignificantOnly | true | Only capture significant content | | minSignificanceScore | 0.5 | Min score to capture (higher = stricter) | | profileFrequency | 15 | Inject profile every N turns (higher = less context) | | maxRecallResults | 5 | Max memories injected per turn | | similarityThreshold | 0.35 | Min relevance to inject | | maxMemoryInjections | 3 | Max memories to show per recall | | contextBudget | 2000 | Max chars of memory context injected | | maxMemories | 500 | Maximum memories to keep | | pruneOlderThanDays | 30 | Auto-delete memories older than N days | | decayRate | 0.05 | Importance decay speed | | importanceWeight | 0.25 | Weight of importance in scoring | | recencyWeight | 0.25 | Weight of recency in scoring | | relevanceWeight | 0.5 | Weight of TF-IDF relevance in scoring |

Data Storage

All memories stored locally in:

~/.openclaw/memory/.json

Default: ~/.openclaw/memory/openclaw_local_memory.json

Privacy

Requirements

Tips

For Best Results

If Context Gets Long

Forgot Something?

{
  "autoRecall": true,
  "autoCapture": true,
  "captureInterval": 8,
  "captureSignificantOnly": true,
  "minSignificanceScore": 0.5,
  "profileFrequency": 15,
  "includeProfileOnFirstTurn": true,
  "maxRecallResults": 5,
  "similarityThreshold": 0.35,
  "maxMemoryInjections": 3,
  "contextBudget": 2000,
  "maxMemories": 500,
  "pruneOlderThanDays": 30,
  "decayRate": 0.05,
  "chunkSize": 800,
  "importanceWeight": 0.25,
  "recencyWeight": 0.25,
  "relevanceWeight": 0.5
}

| Option | Default | Description | |--------|---------|-------------| | autoRecall | true | Inject relevant memories before each turn | | autoCapture | true | Auto-capture conversation exchanges | | captureInterval | 8 | Capture every N turns (higher = less storage) | | captureSignificantOnly | true | Only capture significant content | | minSignificanceScore | 0.5 | Min score to capture (higher = stricter) | | profileFrequency | 15 | Inject profile every N turns (higher = less context) | | maxRecallResults | 5 | Max memories injected per turn | | similarityThreshold | 0.35 | Min relevance to inject | | maxMemoryInjections | 3 | Max memories to show per recall | | contextBudget | 2000 | Max chars of memory context injected | | maxMemories | 500 | Maximum memories to keep | | pruneOlderThanDays | 30 | Auto-delete memories older than N days | | decayRate | 0.05 | Importance decay speed | | importanceWeight | 0.25 | Weight of importance in scoring | | recencyWeight | 0.25 | Weight of recency in scoring | | relevanceWeight | 0.5 | Weight of TF-IDF relevance in scoring |

For Best Results

If Context Gets Long

Forgot Something?