Hybrid search
BM25 + vector + reranking. Six pluggable backends. Find the right memory by meaning, not keyword.
OpenClaw's built-in memory is fine for getting started, but it lacks
semantic search, entity tracking, lifecycle management, and multi-agent
isolation. @remnic/plugin-openclaw is a thin adapter that
replaces the memory slot with the Remnic engine. All data still stays on
your disk as plain markdown.
# 1. Install the plugin package
openclaw plugins install clawhub:@remnic/plugin-openclaw
# 2. Wire up the memory slot automatically
remnic openclaw install
# 3. Restart the gateway
launchctl kickstart -k gui/$(id -u)/ai.openclaw.gateway
# or on Linux:
systemctl restart openclaw-gateway
# 4. Verify
remnic doctor remnic openclaw install writes
plugins.entries["openclaw-remnic"] and
plugins.slots.memory = "openclaw-remnic" to
~/.openclaw/openclaw.json. OpenClaw gates memory plugins
on the slots.memory key; without it, hooks never fire.
openclaw.json, and restart the gateway for you.
After restart, the gateway log should contain:
[remnic] gateway_start fired - Remnic memory plugin is active
id=openclaw-remnic, memoryDir=~/.openclaw/workspace/memory/local On macOS:
grep "gateway_start fired" ~/.openclaw/logs/gateway.log
If the line is missing, run remnic doctor. Each failing
check includes a remediation hint pointing at
remnic openclaw install.
BM25 + vector + reranking. Six pluggable backends. Find the right memory by meaning, not keyword.
Tracks people, projects, tools, and relationships as structured entities. Causal and timeline queries.
Active → validated → stale → archived. Old memories drop out of recall automatically.
Multi-agent isolation with explicit principals. Shared context for agents that should see the same memory.
Review queues, shadow/apply modes, reversible transitions. Trust-zone promotion workflow.
Proactive session archive + hierarchical summary DAG. Context never dies when the window compacts.
{
"plugins": {
"allow": ["openclaw-remnic"],
"slots": { "memory": "openclaw-remnic" },
"entries": {
"openclaw-remnic": {
"enabled": true,
"config": {
// Option 1: OpenAI for extraction
"openaiApiKey": "${OPENAI_API_KEY}"
// Option 2: local LLM (Ollama, LM Studio)
// "localLlmEnabled": true,
// "localLlmUrl": "http://localhost:1234/v1",
// "localLlmModel": "qwen2.5-32b-instruct"
// Option 3: gateway model chain
// "modelSource": "gateway",
// "gatewayAgentId": "remnic-llm",
// "fastGatewayAgentId": "remnic-llm-fast"
}
}
}
}
} The full config reference has 90+ settings covering search backends, capture modes, namespaces, governance, and Memory OS features. See docs/config-reference.md.
With modelSource: "gateway", Remnic routes every LLM call
(extraction, consolidation, reranking) through an OpenClaw agent
persona's model chain instead of its own config. Define personas in
openclaw.json → agents.list[] with a primary
model and fallbacks[] array. Remnic tries each in order
until one succeeds. Build multi-provider chains like
Fireworks → local LLM → cloud OpenAI.
Inside the OpenClaw host, Remnic registers roughly 100 operator
subcommands covering tiering, namespaces, capsules, secure store,
peers, Dreams, governance, and benchmarks. These live under the
openclaw engram command namespace, kept from the former
Engram name during the rename window; see the
rename note. There is no
openclaw remnic shell namespace, and the in-chat session
command is /remnic <on|off|status|clear|stats|flush>.
Full reference:
docs/api.md.