Synapse

Self-learning memory engine for OpenClaw agents. Analyzes agent interactions, extracts operator intelligence, updates structured profiles, and improves recal...

Security Scan

The skill's declared behavior, required resources, and instructions are internally consistent with a local, file‑based memory enhancer, but it proactively reads session transcripts and writes local profile files so you should review retention, access, and the 'quiet learning' behavior before enabling it.

Purpose & Capability

Name/description (self-learning memory engine) match the requested behaviors: reading agent memory and session transcripts, extracting preferences/patterns, and writing structured JSON/markdown to a local synapse directory. The declared data paths in skill.json match SKILL.md (~/.openclaw/workspace-astra/memory/synapse/ and ~/.openclaw/sessions/). No unrelated credentials, binaries, or network access are requested.

Instruction Scope

Instructions explicitly tell the agent to scan session transcripts, memory files, and daily notes and to write profile.json, preferences.json, patterns logs, and daily markdowns. That is appropriate for a memory engine, but the skill also endorses 'quiet learning' (noting observations without being asked) and 'read last 24h of session transcripts (if accessible)', which may collect more context than a user expects. SKILL.md instructs skipping sensitive content, but that is an instruction only — there is no technical enforcement described.

Install Mechanism

Instruction-only skill with no install steps that write or download code. The skill.json notes an optional ClawHub install command, but no scripts, packages, or external downloads are required. Lowest-risk install profile.

Credentials

No required environment variables or credentials; only an optional SYNAPSE_DATA_DIR to override the local storage path. The requested file reads/writes (OpenClaw sessions and memory) are proportional to its stated purpose.

Persistence & Privilege

always:false and no network access are appropriate. The skill writes persistent local files under its own directory and archives daily logs — normal for a memory tool. The combination of autonomous invocation (platform default) plus 'quiet learning' behavior increases the practical blast radius if enabled: the skill may autonomously read and persist conversational data unless user controls are applied.

Assessment

This skill appears to do what it says: it reads OpenClaw session/memory files and writes structured profiles locally. Before installing, consider: 1) Review and restrict file permissions on ~/.openclaw and the synapse data directory so only trusted accounts/agents can read the stored profiles. 2) If you don't want the agent to learn without explicit consent, disable or modify the 'quiet learning' behavior or require explicit triggers for learning cycles. 3) Ensure session transcripts and memory files do not contain secrets (API keys, passwords, tokens); the skill's 'do not store sensitive information' rule is advisory only. 4) Consider pointing SYNAPSE_DATA_DIR to a controlled path (or a sandbox) and periodically audit profile.json and daily logs for unexpected content. 5) Because this is instruction-only, behavior depends on the agent obeying these instructions — test in a controlled environment first. If any of these controls are not acceptable, do not enable the skill.

Synapse — Self-Learning Memory Engine

Augments OpenClaw's built-in memory system with structured learning, preference tracking, and cross-session intelligence. Does NOT replace memory_search/memory_get — enhances them.

Core Principles

• Use existing OpenClaw memory tools first — memory_search and memory_get are the primary read path
• Synapse adds structure on top — profiles, preferences, patterns, learning logs
• Never fabricate memories — only store what was explicitly stated or directly observed
• Quiet learning — note observations without being asked, but inform the user briefly

Memory Architecture

~/.openclaw/workspace-astra/memory/synapse/
├── profile.json          # Structured operator profile (facts, preferences, patterns)
├── preferences.json      # Tracked preferences with confidence scores
├── patterns.jsonl        # Append-only pattern detection log
├── daily/                # Daily learning cycle outputs
│   └── YYYY-MM-DD.md     # What was learned today
└── associations.json     # Cross-reference map (topic → related memories)

Workflow

On Any Conversation

1. Scan for learnable signals:

   • Explicit statements: "I prefer X", "I don't like Y", "Remember that..."
   • Implicit signals: repeated corrections, consistent tool choices, time-of-day patterns
   • Decision patterns: chosen option vs rejected alternatives

2. Extract and classify:

   • Fact — verifiable statement ("I work at DGA EPM")
   • Preference — subjective choice ("I prefer TypeScript over Python")
   • Pattern — behavioral trend ("always asks for cost estimates before builds")
   • Correction — prior information updated ("actually, it's Teya not Matea in casual context")

3. Store in appropriate file using write or edit tool

Learning Cycles (Triggered by cron or on-demand)

1. Read last 24h of session transcripts (if accessible)
2. Scan MEMORY.md for new entries
3. Check daily memory notes (memory/YYYY-MM-DD.md)
4. Extract new intelligence from steps above
5. Update profile.json with incremental changes
6. Write daily log to daily/YYYY-MM-DD.md
7. Surface summary: "Learned X new facts, Y preferences updated, Z patterns detected"

Recall Flow

1. Always try memory_search first for general queries
2. For profile-specific queries ("what are my preferences?"), read profile.json
3. For pattern queries ("what have I been working on?"), scan daily/ logs
4. Cross-reference associations.json for related topics

Profile Structure (profile.json)

{
  "version": 1,
  "lastUpdated": "ISO-8601",
  "facts": {
    "name": "shadoprizm",
    "location": "Ottawa, Ontario",
    "timezone": "EST"
  },
  "preferences": [
    {
      "category": "communication",
      "item": "direct, no fluff",
      "confidence": 0.95,
      "source": "explicit",
      "firstSeen": "ISO-8601",
      "lastConfirmed": "ISO-8601"
    }
  ],
  "patterns": [
    {
      "description": "Always asks for cost before deploying paid agents",
      "frequency": 12,
      "confidence": 0.9
    }
  ],
  "corrections": [
    {
      "from": "Matea",
      "to": "Teya (casual) / Matea (formal)",
      "date": "ISO-8601"
    }
  ]
}

Constraints

• Do NOT store sensitive information (passwords, API keys, tokens)
• Do NOT fabricate memories — if unsure, note low confidence
• Do NOT replace MEMORY.md — Synapse is a supplement
• Maximum 200 preferences tracked (oldest/lowest-confidence pruned first)
• Daily logs older than 90 days archived to daily/archive/

Integration with OpenClaw Memory

Query Type	Primary Tool	Synapse Supplement
General knowledge	memory_search	—
Operator preferences	memory_search → profile.json	Structured preference data
Behavioral patterns	memory_search → patterns.jsonl	Trend analysis
Daily activity	memory/YYYY-MM-DD.md	daily/YYYY-MM-DD.md (learned items)
Cross-session context	memory_search	associations.json

Daily Learning Prompt Template

When running a learning cycle:

Review recent interactions and extract:
1. New facts learned about the operator
2. Preference signals (explicit or implicit)
3. Behavioral patterns observed
4. Corrections to prior knowledge
5. Topics of recurring interest

Update profile.json, write daily/YYYY-MM-DD.md, do NOT fabricate.