Cognition

Practical memory architecture for OpenClaw agents. Helps agents remember commitments, retrieve context, prevent repeat mistakes, compile reusable procedures,...

Security Scan

The skill's files, instructions, and install script are coherent with a memory/memoization architecture and do not request secrets or remote downloads; it's internally consistent but you should review automation behaviors before enabling autonomous execution.

Purpose & Capability

Name/description (memory architecture) matches the manifest: SKILL.md, reference docs, templates, and an install.sh that scaffolds local memory files. No unrelated credentials, binaries, or network endpoints are required.

Instruction Scope

Runtime instructions focus on creating and managing local memory files, staged consolidation, and weekly reports. They require the agent to read/write workspace files (explicitly required by the prompts). One item to watch: the FUTURE_INTENTS template language (‘Execute when triggered’, ‘Scanned every session start + every heartbeat’) implies the agent may act on intents — review how your agent runtime treats 'intents' to avoid unintended automated actions.

Install Mechanism

There is no remote install spec; the included scripts/install.sh is a local scaffold that copies templates into a workspace (default ~/.openclaw/workspace). The script is non-destructive (skips existing files) and contains no downloads or execution of remote code, but it will create files and directories under the chosen workspace path.

Credentials

The skill requests no environment variables, credentials, or config paths. Prompts mention using a 'recommended model' and read/write tools but do not demand secrets or external tokens — proportional to a local memory management skill.

Persistence & Privilege

always:false (no forced permanent inclusion). The skill's workflows expect scheduled jobs (cron) and agent tool access for read/write. Combined with the FUTURE_INTENTS template, this can enable automated behavior if the agent runtime is allowed to autonomously execute actions — consider limiting autonomous execution or reviewing intent-trigger rules.

Assessment

This skill looks coherent for local memory management: it creates a workspace structure, provides consolidation/weekly prompts, and keeps changes staged before mutating durable files. Before installing: 1) Inspect templates (especially FUTURE_INTENTS.md) to confirm you accept any automated 'execute when triggered' semantics. 2) Run scripts/install.sh with an explicit workspace path (not the default) if you want control over where files are created. 3) If you enable nightly/weekly jobs, ensure the scheduled jobs run under an account with only the permissions you intend. 4) Restrict agent tool permissions (read/write/execute) if you do not want the agent to automatically act on FUTURE_INTENTS or perform file mutations without human review. 5) Back up any existing memory files you care about before installing. Overall the skill is internally consistent and does not request secrets or fetch remote code, but be cautious about enabling autonomous execution of intents and scheduled consolidation workflows.

Cognition gives an OpenClaw agent a practical memory architecture: raw logs, durable knowledge, commitments, and reusable procedures.

Use it to make memory operational, not decorative.

Start Here

Install the scaffolding:

bash {baseDir}/scripts/install.sh

Then adopt it in 3 tiers.

Tier 1 — Core

Adopt this first. It is the smallest useful subset.

Core behavior

1. Log important work to memory/YYYY-MM-DD.md
2. Keep durable facts in MEMORY.md and deeper detail in memory/bank/*.md
3. Track commitments in memory/FUTURE_INTENTS.md
4. Add the protocol blocks from {baseDir}/references/protocols.md to your AGENTS.md

What Core gives you

• recent-session continuity
• a compact durable memory index
• fewer dropped promises
• a default place to put important facts instead of hoping the model remembers

Tier 2 — Recommended

Adopt this once Core is part of normal use.

Add these pieces

• nightly staged consolidation via {baseDir}/references/consolidation-prompt.md
• reusable procedures in memory/procedures/
• procedure registry in memory/procedures/index.yaml
• importance tagging from {baseDir}/references/protocols.md

Safety model

• consolidation stages proposals before durable mutation
• daily logs stay append-only
• uncertain items stay tagged [NEEDS_REVIEW]
• new procedures start as draft
• promote procedures cautiously: draft → reviewed → trusted

Tier 3 — Advanced

Only adopt this if Core and Recommended are already working.

Optional overlays

• memory/bank/cross-references.md for linked retrieval
• memory/KNOWLEDGE_MAP.md for coarse confidence tracking
• memory/meta/gap_tracker.json for repeated retrieval misses
• memory/meta/retrieval_log.json for retrieval analytics
• weekly reflection via {baseDir}/references/weekly-reflection-prompt.md
• hybrid retrieval tuning via {baseDir}/references/config.md

Advanced is for sharper retrieval and maintenance, not for replacing the Core system.

Core File Map

Always-on files

• memory/YYYY-MM-DD.md — daily session log
• MEMORY.md — durable fact index
• memory/bank/*.md — deeper topic files
• memory/FUTURE_INTENTS.md — commitments and deferred actions

Recommended files

• memory/consolidation/YYYY-MM-DD-staged.md — staged memory proposals
• memory/procedures/index.yaml — procedure registry
• memory/procedures/*.md — reusable procedures

Advanced files

• memory/KNOWLEDGE_MAP.md
• memory/bank/cross-references.md
• memory/meta/gap_tracker.json
• memory/meta/retrieval_log.json
• memory/summaries/YYYY-WNN.md

Reference Map

Read only what you need:

• {baseDir}/references/protocols.md — protocol blocks to copy into AGENTS.md
• {baseDir}/references/consolidation-prompt.md — nightly staged consolidation
• {baseDir}/references/weekly-reflection-prompt.md — weekly reflection without direct mutation
• {baseDir}/references/architecture.md — core systems, overlays, retrieval tiers, safety notes, file structure
• {baseDir}/references/config.md — advanced hybrid retrieval tuning
• {baseDir}/references/cognitive-science.md — theory appendix

Guardrails

• Prefer staged proposals over silent durable-memory edits
• Keep raw logs append-only
• Keep MEMORY.md compact; move detail into memory/bank/
• Treat procedure promotion as earned, not automatic
• If retrieval fails, log the gap instead of inventing certainty