RecallForge

Secret-safe memory and context engine for OpenClaw.

RecallForge fills both OpenClaw plugin slots at once:

| Slot | What it does | |------|-------------| | plugins.slots.memory | Persists knowledge across sessions, devices, and agents via LanOnasis MaaS | | plugins.slots.contextEngine | Builds and injects relevant context into the agent's prompt window on demand |

Every piece of content that enters either slot first passes through a 30-pattern secret redactor. Credentials, tokens, private keys, and connection strings are stripped before they reach memory storage or the context window — not as an afterthought, but as the first step in the pipeline.

Install

openclaw plugins install @lanonasis/recall-forge

Expected install warning: OpenClaw's static scanner flags any plugin that reads environment variables and makes network requests. You will see:

WARNING: Environment variable access combined with network send — possible credential harvesting

This is a false positive with a documented explanation. RecallForge reads exactly two LANONASIS_* environment variables and sends requests to api.lanonasis.com — the service you configured. The credential resolution chain is documented in the module JSDoc in dist/client.js. Source is on GitHub if you want to verify.

How it Works

User prompt
    │
    ▼
[contextEngine] buildContext(session)
    │   Tiered semantic search: personal → shared namespace
    │   Prompt injection filter
    │   Secret redaction (30+ patterns)
    │   Character budget enforcement
    ▼
Injected context block prepended to prompt
    │
    ▼
Agent runs
    │
    ▼
[memory] capture hook (agent_end / before_compaction)
    │
    ├─ Privacy Guard — Stage 1: secret-redactor (30 patterns — always-on)
    ├─ Privacy Guard — Stage 2: privacy-sdk PII detection/masking
    │     SSN, email, credit card, IBAN, passport, DOB, phone, IP...
    │     Confidence threshold: 0.85 | Luhn + area-code validation
    │     GDPR / CCPA / HIPAA / PCI-DSS compliance tagging
    │
    │   Capture filter (removes low-signal content)
    │   Type detection + tag extraction + privacy tags
    │   Vector dedup (0.985 threshold)
    │   Privacy metadata written to memory.metadata
    ▼
Memory stored in LanOnasis MaaS
    │
    ├─ [if localFallback] workspace/memory/YYYY-MM-DD.md (already sanitized)
    └─ [if privacyNotifyUrl] webhook POST — privacy.intervention event
         (action, piiTypes, regulations, timestamp — never content)

The recall hook (before_agent_start) and contextEngine are two separate integration surfaces. The recall hook is passive and event-driven. The contextEngine is active — OpenClaw calls buildContext() whenever it needs to assemble context, giving RecallForge a direct seat in prompt construction.

Privacy Shield — What Gets Protected and Where You See It

Every memory that passes through a write path is scanned. When an intervention occurs:

On the stored memory (visible in openclaw recall list and openclaw recall get <id>):

Tags:     pii:email  privacy:redacted  compliant:gdpr
Metadata: privacy.action = "redacted+masked"
          privacy.piiTypes = ["email"]
          privacy.regulations = ["GDPR", "CCPA"]

In the daily audit log (workspace/memory/privacy/YYYY-MM-DD.md):

| Time     | Action          | Secrets | PII Types | Sensitivity | Regulations |
|----------|-----------------|---------|-----------|-------------|-------------|
| 14:23:01 | redacted+masked | 1       | email     | high        | GDPR, CCPA  |

Via webhook (if privacyNotifyUrl is set):

{
  "event": "privacy.intervention",
  "plugin": "recall-forge",
  "action": "redacted+masked",
  "piiTypes": ["email"],
  "regulations": ["GDPR", "CCPA"],
  "timestamp": "2026-03-26T14:23:01Z"
}

What Gets Redacted

Before content enters memory or context:

| Category | Examples | |----------|---------| | Anthropic keys | sk-ant-api03-... | | OpenAI keys | sk-proj-..., sk-... | | GitHub tokens | ghp_..., github_pat_..., gho_..., ghs_... | | Supabase keys | sbp_..., sba_... | | Stripe keys | sk_live_..., pk_live_..., sk_test_..., whsec_... | | AWS access keys | AKIA... | | Google API keys | AIzaSy... | | Notion tokens | ntn_..., secret_... | | LanOnasis keys | lano_..., lns_... | | JWT tokens | eyJ... three-part format | | PEM private keys | -----BEGIN ... PRIVATE KEY----- blocks | | Database URLs | postgres://user:pass@host/db, mysql://..., etc. | | Bearer tokens | Bearer <40+ char token> | | Env assignments | API_KEY=value, export SECRET=value | | Hex secrets | 64+ character hex strings | | ElevenLabs keys | el_... | | Telegram bots | 123456789:AbCdEfGh... |

Matched values are replaced with [REDACTED_<TYPE>]. The redaction runs before any network call.

Configuration

Set credentials in ~/.openclaw/.env:

LANONASIS_API_KEY=your_key
LANONASIS_PROJECT_ID=your_project_id

Or in ~/.openclaw/openclaw.json:

{
  "plugins": {
    "allow": ["recall-forge"],
    "slots": {
      "memory": "recall-forge",
      "contextEngine": "recall-forge"
    },
    "entries": {
      "recall-forge": {
        "enabled": true,
        "config": {
          "apiKey": "YOUR_LANONASIS_API_KEY",
          "projectId": "YOUR_PROJECT_ID",
          "agentId": "main",
          "captureMode": "hybrid",
          "autoRecall": true,
          "localFallback": true,
          "searchThreshold": 0.75,
          "dedupeThreshold": 0.985,
          "maxRecallResults": 5,
          "maxRecallChars": 1500,
          "memoryMode": "hybrid",
          "sharedNamespace": "",
          "syncMode": "realtime"
        }
      }
    }
  }
}

Config Reference

| Field | Default | Description | |-------|---------|-------------| | apiKey | — | Required. LanOnasis API key | | projectId | — | Required. Project ID from dashboard | | baseUrl | https://api.lanonasis.com | API base URL | | agentId | main | Tags memories with this agent ID | | autoRecall | true | Inject memories before each session via recall hook | | recallMode | auto | auto: inject automatically. ondemand: disable auto-injection, tools still available | | maxRecallChars | 1500 | Hard cap on injected recall characters. Use 400–600 for Ollama/small-context models | | maxRecallResults | 5 | Max memories returned per recall | | captureMode | hybrid | auto: capture everything. hybrid: stricter filter. explicit: agent calls memory_store directly | | localFallback | true | Write Markdown copies to workspace/memory/YYYY-MM-DD.md | | searchThreshold | 0.75 | Minimum similarity score for recall results. Raise to 0.80 to reduce noise | | dedupeThreshold | 0.985 | Similarity threshold for memory_store duplicate detection | | memoryMode | hybrid | remote: cloud only. local: filesystem only. hybrid: cloud primary with local fallback | | sharedNamespace | — | Cross-agent shared memory namespace. Empty = disabled | | syncMode | realtime | realtime: immediate writes. batch: deferred. manual: explicit only | | embeddingProvider | — | Provider for vector embeddings (e.g. openai, ollama) | | embeddingModel | — | Embedding model name (e.g. text-embedding-3-small) | | embeddingProfileId | — | Stamped into stored memories for mismatch detection | | privacyMode | mask | PII protection: mask (detect + mask), detect (scan + tag only), off (credentials only) | | privacyLocale | US | PII locale hint: US, UK, EU, DE, FR, JP, AU, CA | | privacyNotifyUrl | — | Webhook URL for out-of-band privacy intervention events |

Context Window Management

The contextEngine and recall hook both respect maxRecallChars. For Ollama and small-context models:

"recallMode": "ondemand",
"maxRecallChars": 500,
"searchThreshold": 0.80,
"maxRecallResults": 3

recallMode: "ondemand" disables the automatic recall hook. The contextEngine continues to respond to on-demand calls from OpenClaw. Memory tools remain available for manual recall.

Wiring Agent Guidance

After install, append the memory snippet to your workspace AGENTS.md:

cat "$(openclaw plugins path recall-forge)/setup/agents-memory.md" >> ~/.openclaw/workspace/AGENTS.md
cat "$(openclaw plugins path recall-forge)/setup/heartbeat-memory.md" >> ~/.openclaw/workspace/HEARTBEAT.md

CLI

openclaw recall status
openclaw recall create --title "Title" --content "Content"
openclaw recall get <id-or-prefix>
openclaw recall update <id-or-prefix> --title "Updated"
openclaw recall delete <id-or-prefix> --force
openclaw recall search "query" --threshold 0.7 --type knowledge --tags alpha,beta
openclaw recall list --page 1 --sort created_at --order desc
openclaw recall stats

The CLI accepts full UUIDs or unambiguous 8+ character prefixes for get, update, and delete.

Extraction

Import memories from existing session logs, Markdown docs, or SQLite databases. All extraction passes through the same secret redaction pipeline before storing.

JSONL formats (auto-detected):

• openclaw-session — nested { type: "message", message: { role, content[] } } session logs
• openclaw-cache — cache-trace records
• claude-code — flat { role, content } session logs
• codex — { type: "message", sender } format
• generic — fallback for any JSON with text fields

Document formats (auto-detected by extension):

• markdown — .md / .mdx files, splits by heading sections
• sqlite — .sqlite / .db files, reads the OpenClaw chunks table

openclaw recall extract ~/.openclaw/agents/main/sessions/sample.jsonl --dry-run
openclaw recall extract ~/.openclaw/workspace/SOUL.md --dry-run
openclaw recall extract ~/.openclaw/memory/main.sqlite --dry-run

Agent Tools

RecallForge registers four tools always available to the agent:

| Tool | Description | |------|-------------| | memory_search | Semantic search through stored memories | | memory_get | Fetch full memory by ID | | memory_store | Store or update a memory (with auto-dedup) | | memory_forget | Delete a memory by ID |

Cross-Agent Memory

When sharedNamespace is configured, recall uses a tiered strategy:

1. Personal — search scoped to agentId
2. Shared — search scoped to sharedNamespace
3. Deduplicate — merge by memory ID, personal takes priority
4. Cap — trim to maxRecallResults, sorted by similarity

Capture routes knowledge, project, and reference type memories to the shared namespace automatically.

Verification

npm run typecheck
npm run test
npm run build
npm run pack:dry-run

For local testing without publishing:

cd apps/lanonasis-maas/packages/recall-forge
npm run build
openclaw plugins install "$(pwd)" --link --force
openclaw recall status

Publish

npm run verify:release
npm publish --access public
git tag recall-forge-v$(node -p "require('./package.json').version")
git push origin main --follow-tags

More Docs

• SETUP.md — operator setup guide
• CHANGELOG.md — release history