🧠 Cognition — Seven Memory Systems for OpenClaw Agents

Give your agent a mind, not just memory.

What It Does

Your agent forgets everything between sessions. Cognition fixes that with seven cognitive systems borrowed from neuroscience:

1. Working Memory — Context window management with GWT broadcasting
2. Episodic Memory — Session replay with importance tagging
3. Semantic Memory — Hierarchical knowledge with activation-weighted retrieval
4. Procedural Memory — Compiled skills with Bayesian success tracking
5. Prospective Memory — Structured future intents (never forget a commitment)
6. Metamemory — Confidence scoring and knowledge gap detection
7. Causal-Temporal Reasoning — Cross-references and spreading activation

Quick Start

Run the install script to scaffold the full directory structure:

bash {baseDir}/scripts/install.sh

This creates all template files and directories. Then follow the 4-phase adoption guide below.

Phase 1: Foundation (10 minutes)

1.1 Add FUTURE_INTENTS.md to workspace root

Copy from {baseDir}/templates/FUTURE_INTENTS.md. This is your agent's prospective memory — structured commitments with triggers, actions, and status tracking.

1.2 Create procedural memory

The install script creates memory/procedures/ with index.yaml. After solving non-trivial problems, compile solutions here. Each procedure tracks preconditions, steps, failure modes, and success rate.

1.3 Set up nightly consolidation

Add a cron job using the prompt at {baseDir}/references/consolidation-prompt.md. Schedule: 0 2 * * * (2 AM daily). Uses your agent's model to process the day's logs into durable knowledge.

Recommended model: Any model with reliable tool use (Sonnet 4+, GPT-5+). Small local models may narrate instead of executing — test tool reliability first.

1.4 Add Memory Protocol to AGENTS.md

Paste the protocol block from {baseDir}/references/protocols.md into your AGENTS.md.

Phase 2: Consolidation Intelligence (10 minutes)

2.1 Add KNOWLEDGE_MAP.md

Copy from {baseDir}/templates/KNOWLEDGE_MAP.md. Customize domains for your use case. Confidence scores: 🟢 High (0.8+) | 🟡 Medium (0.5-0.8) | 🔴 Low (<0.5).

2.2 Enable importance tagging

Add the tagging protocol from {baseDir}/references/protocols.md to AGENTS.md. Tag daily log entries:

• [REPLAY_PRIORITY: HIGH] — Corrections, policy changes, decisions
• [REPLAY_PRIORITY: MEDIUM] — New facts, milestones, config changes
• LOW = default, no tag needed

2.3 Enable GWT Broadcasting

Add the broadcasting rules from {baseDir}/references/protocols.md. When important info enters a session, push it to the correct store immediately.

Phase 3: Intelligence (15 minutes)

3.1 Create cross-references

Copy {baseDir}/templates/cross-references.md to memory/bank/. Customize with your person→file, project→resource, and domain→domain links. This enables spreading activation — retrieving one topic primes related topics.

3.2 Set up gap tracking

The install script creates memory/meta/gap_tracker.json. Failed searches are logged here. Gaps with 3+ misses are surfaced during weekly reflection.

3.3 Upgrade weekly reflection

Add a weekly cron using the prompt at {baseDir}/references/weekly-reflection-prompt.md. Schedule: 0 4 * * 0 (Sunday 4 AM). Performs 13-step deep cognitive maintenance.

3.4 Update Retrieval Protocol

Replace your basic retrieval protocol with the enhanced version from {baseDir}/references/protocols.md — adds cross-reference following and gap logging.

Phase 4: Evolution (ongoing)

Phase 4 is emergent. As the other systems accumulate data:

• Weekly reflection recommends AGENTS.md rule changes
• Procedures compile automatically from solved problems
• Knowledge gaps drive proactive research
• Confidence scores guide when to act vs. when to ask

Configuration

Recommended openclaw.json settings

Read {baseDir}/references/config.md for copy-paste config blocks:

• reserveTokensFloor: 40000 (not the default 20K)
• memoryFlush: enabled with 4K-10K soft threshold
• Hybrid search: 0.7 vector / 0.3 BM25
• MMR diversity: lambda 0.7
• Temporal decay: halfLifeDays 30
• Embedding cache: 50K entries

File Structure

After installation:

workspace/
├── FUTURE_INTENTS.md          # Prospective memory
├── KNOWLEDGE_MAP.md           # Metamemory
└── memory/
    ├── bank/
    │   └── cross-references.md  # Spreading activation
    ├── meta/
    │   ├── gap_tracker.json     # Knowledge gap detection
    │   └── retrieval_log.json   # Search quality tracking
    ├── procedures/
    │   └── index.yaml           # Compiled skills registry
    ├── consolidation/
    │   └── YYYY-MM-DD.md        # Nightly consolidation logs
    └── summaries/
        └── YYYY-WNN.md          # Weekly reflection summaries

Cognitive Science Foundation

Theory	Mechanism	System
SOAR	Procedural chunking	Procedural Memory
ACT-R	Activation-weighted retrieval	Semantic Memory
Global Workspace Theory	Broadcast to all stores	GWT Broadcasting
Hippocampal Replay	Overnight consolidation	Nightly Cron
Predictive Processing	Importance tagging	Episodic Memory
Metacognition	Confidence calibration	Metamemory