Summary

Add a cognitive memory layer based on the Atkinson-Shiffrin model (1968) that gives OpenClaw persistent, semantic memory across sessions — with natural forgetting curves, metacognitive error prevention, and trust-based relationship tracking.

Problem to solve

OpenClaw sessions are stateless. Every new conversation starts from zero:

• No memory of past decisions — the assistant doesn't know what was decided last week or why
• Repeats mistakes — makes the same error the user corrected yesterday because it has no learning system
• No context continuity — can't connect today's work with last week's patterns
• Keyword-only search — if you don't remember the exact word, you can't find the memory
• No relationship — treats every session identically regardless of history with the user

This is the #1 frustration for power users who work with AI assistants daily.

Proposed solution

NEXO Brain — an open-source cognitive memory system already implemented and published:

• npm: npx nexo-brain (one command install)
• GitHub: https://github.com/wazionapps/nexo
• License: MIT

What it provides:

Memory Model (Atkinson-Shiffrin):

• Sensory Register → Short-Term Memory (7-day half-life) → Long-Term Memory (60-day half-life)
• Ebbinghaus forgetting curves — memories fade without use, strengthen with access
• Automatic consolidation: frequently-used STM promotes to LTM
• Dormant memories reactivate when relevant context appears

Semantic Search (Vector RAG):

• fastembed (BAAI/bge-small-en-v1.5, 384 dims, CPU-only, no API needed)
• Finds memories by meaning, not just keywords
• "deploy problems" finds "SSH connection timeout on production" even with zero word overlap

Metacognitive Guard:

• Before every code change: "Have I made this mistake before?"
• Searches past errors semantically, surfaces warnings BEFORE acting
• Trust score modulates verification rigor (low trust = more checks)

Cognitive Dissonance Protocol:

• Detects when new instructions conflict with established knowledge
• Verbalizes the conflict instead of silently obeying or resisting
• Three resolutions: paradigm_shift, exception, override

Discriminative Fusion:

• Near-duplicate memories for different contexts (Linux vs Mac) stay separate as "siblings"
• Prevents "lobotomy by averaging" when merging similar memories

Trust Score (0-100):

• Tracks alignment with the user over time
• Positive events (thanks, successful delegation) increase trust
• Errors and corrections decrease trust
• Score modulates internal rigor, not permissions

Sentiment Detection:

• Adapts tone based on user's mood (concise when frustrated, proactive when positive)

Automated Sleep Cycle:

• 03:00 — Ebbinghaus decay + memory consolidation + duplicate merging
• 07:00 — Self-audit + health checks + metrics
• 23:30 — Post-mortem + pattern extraction from day's events
• Boot — Catch-up for any missed processes

Technical specs:

• 50+ MCP tools across 12 categories
• Hot-loadable plugin system
• SQLite-only (no external DB)
• ~100KB package, ~200MB with dependencies (fastembed+numpy)
• macOS (LaunchAgents), Linux support planned

Alternatives considered

1. Simple key-value memory — Doesn't support semantic search or forgetting. Accumulates noise indefinitely.
2. Cloud-based vector DB (Pinecone, Weaviate) — Adds latency, cost, and privacy concerns. NEXO runs 100% local.
3. LLM-summarized memory — Expensive (API calls for every memory operation) and lossy (summaries discard detail).
4. No memory (current state) — Users compensate with manual CLAUDE.md/AGENTS.md maintenance. Doesn't scale.

Impact

• Affected: Every user who works with OpenClaw across multiple sessions
• Severity: High — memory loss is the fundamental limitation of current AI assistants
• Frequency: Every session start (100% of users experience the cold-start problem)
• Consequence: Repeated mistakes, lost context, manual documentation overhead, reduced trust in the assistant

Evidence/examples

• Published and working: https://github.com/wazionapps/nexo
• npm package: nexo-brain (v0.1.1)
• Based on established cognitive psychology models (Atkinson-Shiffrin 1968, Ebbinghaus 1885)
• In production use since March 2026 with 530+ LTM vectors, 60+ STM vectors
• Retrieval relevance: 100% (all queries return useful results above 0.6 cosine threshold)

Additional information

This could integrate with OpenClaw as:

1. A ClawHub skill — standalone extension users can install
2. A core memory module — native integration into OpenClaw's architecture
3. Both — skill for immediate use, native integration as a roadmap item

The entire system is MIT-licensed and designed to be framework-agnostic. The MCP tools can be adapted to OpenClaw's extension API.

We're happy to contribute directly, adapt the architecture to OpenClaw's patterns, or collaborate on integration design.