Agent Memory Setup

Agent Memory Setup

Set up a complete 3-tier memory system for any OpenClaw agent. Includes directory structure, memory files, semantic search, and context compaction.

Quick Start

# 1. Run setup script
bash scripts/setup_memory.sh /path/to/workspace
2. Copy AGENTS.md template to workspace
(read references/AGENTS_TEMPLATE.md, adapt, write to workspace/AGENTS.md)
3. Add config to openclaw.json (see Step 3 below for exact JSON)
4. Restart
openclaw gateway restart

For full details, read the sections below.

When NOT to Use This Skill

Prerequisites

Before running setup, ensure:

Platform Notes

Setup Steps

Step 1: Run the setup script

bash scripts/setup_memory.sh /path/to/agent/workspace

This creates:

It also checks for QMD and Lossless Claw, installing them if possible.

Step 2: Copy the AGENTS.md template

Read references/AGENTS_TEMPLATE.md and write it to the agent's workspace as AGENTS.md.

What to customize:

What to keep as-is:

Step 3: Configure openclaw.json

Add to agents.defaults (or the specific agent config):

"memorySearch": { "provider": "local" },
"compaction": { "mode": "safeguard" },
"contextPruning": { "mode": "cache-ttl", "ttl": "1h" },
"heartbeat": { "every": "1h" }

What each setting does:

| Setting | Purpose | Customization | |---------|---------|---------------| | memorySearch | Enables QMD semantic search over memory files | "local" is the only current provider | | compaction: safeguard | Lossless Claw compacts old messages into expandable summaries instead of dropping them. "Safeguard" mode triggers compaction before context overflows, preserving everything via lcm_expand | Change mode only if you understand Lossless Claw internals | | contextPruning: cache-ttl | Evicts stale context from the active window after the TTL expires. Works WITH Lossless Claw: content is compacted first, then pruned, so nothing is truly lost | Adjust ttl to match your use case: "5m" for fast-cycling agents, "2h" for long research sessions | | heartbeat | Triggers periodic check-ins where the agent reads HEARTBEAT.md and performs maintenance tasks (email checks, memory review, etc.) | Adjust "every" interval: "5m" for monitoring agents, "1h" for general use, "4h" for low-activity agents |

Enable these plugins for the agent:

"session-memory": { "enabled": true },
"bootstrap-extra-files": { "enabled": true },
"lossless-claw": { "enabled": true }

What each plugin does:

| Plugin | Role | |--------|------| | session-memory | Persists and loads memory context across sessions. Ensures the agent reads memory files (HOT/WARM/COLD) on startup | | bootstrap-extra-files | Loads additional workspace files (AGENTS.md, SOUL.md, USER.md, TOOLS.md) into the agent's context at session start | | lossless-claw | Compacts old conversation into summaries that can be expanded back on demand via lcm_expand, lcm_grep, and lcm_expand_query. Prevents amnesia in long conversations |

Manual plugin install (if the setup script didn't install them):

pip install qmd                # or: pipx install qmd / brew install qmd
openclaw plugins install @martian-engineering/lossless-claw

Step 4: Restart and verify

openclaw gateway restart

Verification checklist — run each and confirm:

# 1. Memory directories exist
ls -d memory/ memory/hot/ memory/warm/
2. Memory files exist
ls memory/hot/HOT_MEMORY.md memory/warm/WARM_MEMORY.md MEMORY.md memory/heartbeat-state.json
3. Today's daily log exists
ls memory/$(date +%Y-%m-%d).md
4. QMD is installed
qmd --version
5. Lossless Claw plugin is active
openclaw plugins list | grep lossless-claw
6. AGENTS.md is in place
head -5 AGENTS.md
7. Config is applied (check openclaw.json)
grep -c "memorySearch\|compaction\|contextPruning\|lossless-claw" ~/.openclaw/openclaw.json

If any check fails, see the Troubleshooting section below.

Migrating Existing Memory

If you already have MEMORY.md or daily logs from before this system:

1. Run the setup script normally — it checks if [ ! -f ] before creating each file, so your existing files are preserved untouched. 2. Reorganize existing content into tiers: - Move active/temporary context from MEMORY.md → memory/hot/HOT_MEMORY.md - Move stable preferences, API refs, recurring config → memory/warm/WARM_MEMORY.md - Keep long-term decisions, milestones, lessons learned in MEMORY.md (COLD tier) 3. Existing daily logs stay in memory/ as-is — the system reads them from there already.

Recovery

Corrupted memory file: Delete the corrupted file and re-run bash scripts/setup_memory.sh /path/to/workspace. The script only creates files that don't exist, so other memory files are safe. Alternatively, manually recreate the file with the template header from the setup script.

Lost or deleted AGENTS.md: Re-copy from references/AGENTS_TEMPLATE.md and customize.

Broken openclaw.json: Restore from backup (the setup script doesn't modify openclaw.json — config changes are manual). If no backup exists, re-add the config from Step 3 above.

General principle: The setup script is always safe to re-run. It never overwrites existing files.

Using Lossless Claw After Setup

Once installed, Lossless Claw works automatically:

1. Compaction is automatic — as conversations grow long, old messages are compacted into summaries behind the scenes. 2. The agent retrieves context on demand using these tools: - lcm_expand — expand a compacted summary back into its original messages - lcm_grep — search across all compacted history by regex or text - lcm_expand_query — ask a question and get answers from compacted context 3. No agent code changes needed — if the agent has access to lcm_* tools (standard in OpenClaw), it can retrieve anything that was compacted. 4. AGENTS.md instructs the agent to update HOT_MEMORY before compaction flushes, so critical working state survives even when context is pruned.

How the Tiers Work

| Tier | File | Purpose | Update Frequency | |------|------|---------|-----------------| | 🔥 HOT | memory/hot/HOT_MEMORY.md | Current task, pending actions | Every few turns | | 🌡️ WARM | memory/warm/WARM_MEMORY.md | Stable preferences, API refs, gotchas | When things change | | ❄️ COLD | MEMORY.md | Milestones, decisions, distilled lessons | Weekly/monthly |

Concrete examples per tier

🔥 HOT — "Currently debugging auth flow for client X. Waiting on API key from Bob. Next step: test endpoint once key arrives."

🌡️ WARM — "User prefers bullet lists over paragraphs. Slack workspace is #eng-team. API rate limit is 100 req/min. Always use UTC timestamps."

❄️ COLD — "2025-03-15: Migrated from v1 to v2 API. Key lesson: always pin dependency versions. Decision: adopted trunk-based development."

Daily logs (memory/YYYY-MM-DD.md) capture raw session events. Periodically, the agent reviews daily logs and promotes important items up to COLD.

Multi-Agent & Multi-Context Workspaces

Shared workspaces (multiple agents, one directory)

Separate contexts (one agent, multiple domains)

Each OpenClaw agent maps to one workspace. If you need separate memory contexts (e.g., work vs. personal):

Long-Term Maintenance

Daily logs grow over time. Each day creates a new memory/YYYY-MM-DD.md file. Over a year, this accumulates ~365 small files. This is generally fine (each file is small — typically 1–10 KB), but for long-running agents:

Troubleshooting

| Symptom | Likely Cause | Fix | |---------|-------------|-----| | Agent doesn't read memory files on startup | AGENTS.md missing or doesn't include memory instructions | Copy references/AGENTS_TEMPLATE.md → AGENTS.md | | memory_search not working | QMD not installed or memorySearch not configured | Run qmd --version; add "memorySearch": { "provider": "local" } to config | | Old conversations vanish (no summaries) | Lossless Claw not enabled | Check openclaw plugins list for lossless-claw; enable in plugins config | | Config changes have no effect | Gateway not restarted | Run openclaw gateway restart | | Plugin install fails | openclaw CLI not in PATH or npm issue | Verify openclaw --version; try npm install -g openclaw | | Agent overwrites existing memory files | Script bug (shouldn't happen) | Script checks if [ ! -f ] before creating — report if override occurs |

# 1. Run setup script
bash scripts/setup_memory.sh /path/to/workspace
2. Copy AGENTS.md template to workspace
(read references/AGENTS_TEMPLATE.md, adapt, write to workspace/AGENTS.md)
3. Add config to openclaw.json (see Step 3 below for exact JSON)
4. Restart
openclaw gateway restart

For full details, read the sections below.

Before running setup, ensure:

Platform Notes

| Symptom | Likely Cause | Fix | |---------|-------------|-----| | Agent doesn't read memory files on startup | AGENTS.md missing or doesn't include memory instructions | Copy references/AGENTS_TEMPLATE.md → AGENTS.md | | memory_search not working | QMD not installed or memorySearch not configured | Run qmd --version; add "memorySearch": { "provider": "local" } to config | | Old conversations vanish (no summaries) | Lossless Claw not enabled | Check openclaw plugins list for lossless-claw; enable in plugins config | | Config changes have no effect | Gateway not restarted | Run openclaw gateway restart | | Plugin install fails | openclaw CLI not in PATH or npm issue | Verify openclaw --version; try npm install -g openclaw | | Agent overwrites existing memory files | Script bug (shouldn't happen) | Script checks if [ ! -f ] before creating — report if override occurs |