Configuration

HindClaw uses a two-level config system: plugin-level defaults in openclaw.json and per-agent overrides in bank config files.

openclaw.json (plugin config)          banks/agent-1.json5 (bank config)
├── Daemon (global only)               ├── Server-side (agent-only)
│   apiPort, embedVersion              │   retain_mission, entity_labels
│   embedPackagePath, daemonIdleTimeout│   dispositions, directives
│                                      │
├── Defaults (overridable per-agent)   ├── Infrastructure overrides
│   hindsightApiUrl, hindsightApiToken │   hindsightApiUrl, hindsightApiToken
│   llmProvider, llmModel              │
│   autoRecall, autoRetain, ...        ├── Behavioral overrides
│                                      │   recallBudget, retainTags, llmModel
├── bootstrap: true|false              │
│                                      ├── Multi-bank: recallFrom [...]
└── Agent mapping                      ├── Session start: sessionStartModels
    agents: { id: { bankConfig } }     └── Reflect: reflectOnRecall

Resolution: plugin defaults -> bank config file (shallow merge, bank file wins).

Plugin Config Reference

Option	Default	Per-agent	Description
hindsightApiUrl	-	yes	Hindsight API URL
hindsightApiToken	-	yes	Bearer token for API auth
apiPort	9077	no	Port for local daemon
embedVersion	"latest"	no	hindsight-embed version
embedPackagePath	-	no	Local hindsight-embed path (development)
daemonIdleTimeout	0	no	Daemon idle timeout in seconds (0 = never)
dynamicBankId	true	yes	Derive bank ID from context
dynamicBankGranularity	["agent","channel","user"]	yes	Fields for bank ID derivation
bankIdPrefix	-	yes	Prefix for derived bank IDs
autoRecall	true	yes	Inject memories before each turn
autoRetain	true	yes	Retain conversations after each turn
recallBudget	"mid"	yes	Recall effort: low, mid, high
recallMaxTokens	1024	yes	Max tokens injected per turn
recallTypes	["world","experience"]	yes	Memory types to recall
retainRoles	["user","assistant"]	yes	Roles captured for retention
retainEveryNTurns	1	yes	Retain every Nth turn
llmProvider	auto	yes	LLM provider for extraction
llmModel	provider default	yes	Model name
bootstrap	false	no	Auto-apply bank configs on first run
agents	{}	no	Per-agent bank config registration

Multi-Server Support

Per-agent infrastructure overrides enable connecting different agents to different Hindsight servers:

Gateway
├── agent-1  (private)  -> hindsightApiUrl: "https://hindsight.home.local"
├── agent-2  (private)  -> hindsightApiUrl: "https://hindsight.home.local"
├── agent-3  (company)  -> hindsightApiUrl: "https://hindsight.office.local"
├── agent-4  (company)  -> hindsightApiUrl: "https://hindsight.office.local"
└── agent-5  (local)    -> no hindsightApiUrl (local daemon)

Remote Hindsight Server

To use a remote/shared Hindsight server instead of the local daemon:

{
  "hindsightApiUrl": "https://hindsight.office.local",
  "hindsightApiToken": "your-api-token"
}

Cross-Agent Recall

An agent can recall from multiple banks using recallFrom in its bank config:

{
  "recallFrom": ["agent-1", "agent-2", "agent-3"],
  "recallBudget": "high",
  "recallMaxTokens": 2048
}

When access control is active, permissions are checked independently for each target bank. If the user has recall: false on a target bank, that bank is skipped entirely.