Memoria

Human-like persistent memory for AI agents. 21 cognitive layers, knowledge graph, procedural learning, continuous capture, bring your own LLM (Ollama, LM Studio, or API), 100% local-first on SQLite, zero cloud cost.

openclaw plugins install clawhub:memoria-plugin

{
  "builtWithOpenClawVersion": "3.25.1",
  "pluginApiRange": ">=1"
}

{
  "bundledSkills": [],
  "capabilityTags": [
    "executes-code"
  ],
  "channels": [],
  "commandNames": [],
  "configSchema": true,
  "configUiHints": false,
  "executesCode": true,
  "hooks": [],
  "httpRouteCount": 0,
  "materializesDependencies": false,
  "providers": [],
  "runtimeId": "memoria",
  "serviceNames": [],
  "setupEntry": false,
  "toolNames": []
}

{
  "hasProvenance": false,
  "scanStatus": "clean",
  "scope": "artifact-only",
  "sourceCommit": "ae242ccba2cb0f54f69b260d9075b28cc742606d",
  "sourceRepo": "Primo-Studio/openclaw-memoria",
  "sourceTag": "ae242ccba2cb0f54f69b260d9075b28cc742606d",
  "summary": "Validated package structure and linked the release to source metadata.",
  "tier": "source-linked"
}

{
  "latest": "3.25.1",
  "memory": "3.4.1",
  "plugin": "3.4.1"
}

🧠 Memoria — Persistent Memory for OpenClaw

Memory that thinks like you do. Your AI assistant remembers what matters, forgets what doesn't, and gets better over time.

SQLite-backed · Fully local · Zero cloud · 21 memory layers · Human-like architecture

✨ What's New in v3.22.3

🔄 Continuous Learning — Layer 21 (v3.22.0)

Memoria no longer waits for end-of-session to learn. New real-time capture via message_received + llm_output hooks:

• 3 extraction modes: periodic (every N turns), urgent (on user frustration/error), self-error (on assistant self-admission)
• Cross-layer integration: extracted facts go through the full pipeline (selective dedup → embed → graph → topics → observations → clusters → sync)
• Smart dedup with agent_end: avoids double LLM calls when continuous already captured
• 6 bugs fixed across 3 audit rounds (v3.22.0 → v3.22.1 → v3.22.3)
• Node 24.x compatibility — fixed CONTINUOUS_ENABLED TDZ crash on gateways with embedded Node 24.x

🔍 Deep Audit — 10+6 bugs found & fixed (v3.21.0–v3.22.3)

Full code audit revealed critical alignment issues:

• Hebbian learning was 100% dead — wrong column names since creation
• Proactive revision never triggered — searched for obsolete lifecycle state
• storeFact() lost 6 columns on INSERT
• Concurrent extraction risk and buffer never cleared in continuous learning
• All 21 layers now properly aligned with the actual database schema

🧩 Behavioral Patterns (v3.19.0)

Detects repeated similar facts and consolidates them into patterns.

🔗 Cross-Layer Connections (v3.20.0)

• Feedback → Lifecycle: facts recalled 5+ times with positive usefulness → auto-promoted to "settled"
• Hebbian → Topics: strong entity relations auto-organize topics into parent/child hierarchy
• Lifecycle → Patterns: confirmed patterns (5+ occurrences) → settled (never forgotten)

✨ Core Features

• 21 memory layers — from text search to procedural memory, knowledge graphs, behavioral patterns, continuous learning, and cross-layer connections
• Semantic vs Episodic — durable knowledge decays slowly, dated events fade (like human memory)
• Lifecycle management — facts evolve: fresh → settled → dormant (not "mature/archived")
• Observations — living syntheses that evolve as new evidence appears
• Fact Clusters — entity-grouped summaries with tracked membership
• Procedural memory — captures "how to do things" with steps, gotchas, quality scores, and failure reasons
• Behavioral patterns — detects repeated preferences and consolidates them
• Adaptive recall — injects 2-12 facts based on context load
• Hot Tier — frequently accessed facts (5+ recalls) always recalled
• Feedback loop — usefulness/recall_count/used_count tracked per fact
• Cross-layer connections — feedback→lifecycle, hebbian→topics, lifecycle→patterns
• Hebbian reinforcement — knowledge graph relations strengthen on co-occurrence, decay when unused
• Proactive revision — settled facts get LLM review and potential update
• Identity-aware — parses SOUL.md/USER.md to prioritize relevant facts
• Expertise specialization — topic access frequency boosts recall
• Provider-agnostic — Ollama, LM Studio, OpenAI, OpenRouter, Anthropic
• Fallback chain — if primary LLM fails, next one takes over; if all fail, facts still stored
• Zero config — smart defaults, 60-second setup

🚀 Quick Install

As Plugin (recommended)

openclaw plugins install memoria-plugin

From Source (review code first)

cd ~/.openclaw/extensions
git clone https://github.com/Primo-Studio/openclaw-memoria.git memoria
cd memoria && npm install

💡 Configure after install via bash ~/.openclaw/extensions/memoria/configure.sh

Minimal manual config

Add to openclaw.json:

{
  "plugins": {
    "allow": ["memoria"],
    "entries": {
      "memoria": { "enabled": true }
    }
  }
}

Smart defaults: Ollama + gemma3:4b + nomic-embed-text-v2-moe (local, 0€).

See INSTALL.md for advanced options.

🏗️ Architecture — 21 Layers

┌──────────────────────────────────────────────────────┐
│                   MEMORIA v3.22.3                    │
├──────────────────────────────────────────────────────┤
│                                                      │
│  RECALL (before each response):                      │
│  Budget → Observations → Hybrid Search (FTS5+embed)  │
│  → Knowledge Graph → Topics → Context Tree           │
│  → Lifecycle mult × Expertise boost × Cluster penalty│
│  → Format + Inject                                   │
│                                                      │
│  CAPTURE (after each conversation):                  │
│  LLM Extract → Selective (dedup/contradiction)       │
│  → Embed → Graph → Hebbian → Topics → Observations  │
│  → Clusters → Patterns → Cross-layer connections     │
│  → Sync .md → Auto-regen                             │
│                                                      │
│  CONTINUOUS (real-time, during conversation):         │
│  message_received + llm_output → rolling buffer      │
│  → periodic/urgent/self-error triggers               │
│  → LLM Extract → same pipeline as CAPTURE            │
│                                                      │
│  LEARNING (background):                              │
│  Feedback (usefulness/recall/used) → Lifecycle       │
│  Hebbian (co-occurrence) → Topic hierarchy           │
│  Pattern detection → Consolidation                   │
│  Proactive revision → Fact evolution                 │
│                                                      │
├──────────────────────────────────────────────────────┤
│  SQLite (FTS5 + vectors) · No cloud required         │
└──────────────────────────────────────────────────────┘

Layer Map

9 layers use LLM via the Fallback Chain. 12 layers are pure algorithmic.

For scoring formulas, decay rates, and detailed pipeline descriptions, see docs/ARCHITECTURE.md.

⚙️ Configuration

{
  "autoRecall": true,
  "autoCapture": true,
  "recallLimit": 12,
  "captureMaxFacts": 8,
  "syncMd": true,

  "llm": {
    "provider": "ollama",
    "model": "gemma3:4b"
  },

  "embed": {
    "provider": "ollama",
    "model": "nomic-embed-text-v2-moe",
    "dimensions": 768
  },

  "fallback": [
    { "provider": "ollama", "model": "gemma3:4b" },
    { "provider": "lmstudio", "model": "auto" }
  ]
}

Supported Providers

📊 Benchmarks

Tested on LongMemEval-S (30 questions, 5 categories):

v3.14–3.21 benchmarks pending (Sol benchmark planned).

Detailed methodology and scripts in benchmarks/.

🗺️ Roadmap

📄 License

Apache License 2.0 — see LICENSE.

Copyright 2026 Primo-Studio by Neto Pompeu.