Cogmem Memory Backend

cogmem — bio-inspired agent memory kernel for OpenClaw. Stores conversation turns, enables recall from dozens of days ago. Requires embedding + LLM model. Single-agent only, not a knowledge base.

cogmem Memory Backend for OpenClaw

Installs and connects the exact version of cogmem from GitHub — agent-native memory kernel — to the current OpenClaw workspace.

This skill follows the official OpenClaw backend documentation from the repository.

What This Does

1. Installs cogmem@2.0.2 CLI globally via the official one-line installer
2. Runs cogmem init with project scope for OpenClaw agent
3. Runs cogmem doctor --fix --agent openclaw --workspace . to validate and repair
4. Runs cogmem connect openclaw --workspace . --auto --force to wire OpenClaw plugin and import existing memory
5. Sets up automatic memory injection and turn recording

⚠️ Requirements

Two models are required:

• Embedding model — for semantic recall (e.g. Ollama qwen3-embedding:0.6b, dimension 1024)
• LLM model — for Dream Curator / memory synthesis (e.g. Ollama qwen2.5:7b)

Without both configured, cogmem doctor will warn and recall quality will be degraded.

What cogmem Is (and Is Not)

✅ It stores: conversation turns, user preferences, agent observations, task events

❌ It is not: a knowledge base, a note-taking app, a vector RAG wrapper, an Obsidian replacement, a shared memory system

Bio-inspired architecture: raw evidence is preserved; active memory is selective. Enables recall from dozens of days ago with source anchors. Each agent has its own isolated memory database.

Install

From the OpenClaw workspace root (~/.openclaw/workspace):

# Install exact version matching the repo
curl -fsSL https://raw.githubusercontent.com/liuqin164/cogmem/main/install.sh | bash
cogmem init --yes --agent openclaw --scope project
cogmem doctor --fix --agent openclaw --workspace .
cogmem connect openclaw --workspace . --auto --force

Local Quantized Embeddings (Recommended)

For semantic recall, configure a local embedding provider. Recommended with Ollama:

ollama pull qwen3-embedding:0.6b

Edit ~/.cogmem/config.toml (or project .cogmem/config.toml):

[core]
db_path = "memory.db"
vector_backend = "sqlite-vec"
vector_dimension = 1024

[embedding]
provider = "openai_compatible"
base_url = "http://localhost:11434/v1"
model = "qwen3-embedding:0.6b"
timeout_ms = 30000

[memory_model]
provider = "openai_compatible"
base_url = "http://localhost:11434/v1"
model = "qwen2.5:7b"
api_key = ""
timeout_ms = 60000

Run cogmem doctor --fix --agent openclaw --workspace . after editing.

Import Existing OpenClaw Memory

Preview first, then migrate:

cogmem import-openclaw --workspace . --project openclaw --dry-run
cogmem import-openclaw --workspace . --project openclaw

For older imports needing raw ledger anchors:

cogmem import-openclaw --workspace . --project openclaw --config .cogmem/config.toml --reindex-raw --json

Update

cogmem update --yes
cogmem doctor --fix --agent openclaw --workspace .

Runtime

After installation, OpenClaw will automatically:

• Inject relevant memories before each prompt via governed recall
• Record each turn (user + assistant) to raw ledger
• Run Dream Curator in background to propose long-term memories
• Use CPU governance to promote evidence-backed summaries

Agent-Facing Recall

You can manually query memories:

cogmem memory recall --query "what did we discuss about memory black boxes?" --project openclaw --agent openclaw --json

Special intents:

• --intent previous_session_summary for "上个会话我们聊了什么"
• --intent forensic_quote for "我当时关于X的原话是什么"

View source context:

cogmem memory show --event <event-id> --before 2 --after 2 --json

Post-Install Verification

cogmem doctor --agent openclaw --workspace .
cogmem memory status --project openclaw --json