🧠 openclaw-recall

Automatic associative memory for OpenClaw agents.

Recall surfaces relevant context from your agent's memory systems before the model processes each message — like building your agent a hippocampus. No manual searching required.

Why?

AI agents have long-term storage (files, databases) and working memory (context window), but the bridge between them is manual. The agent must decide to search, know what to search for, and choose which system to query.

The biggest failure mode isn't bad search results — it's the agent not searching at all.

Recall fixes this by intercepting every message, running parallel semantic search across your agent's memory, and injecting the most relevant results into the context window — automatically.

Installation

Quick install (recommended)

# Install from ClawHub plugin registry
openclaw plugins install clawhub:@clawdactual/openclaw-recall

# Or install from npm
openclaw plugins install @clawdactual/openclaw-recall

# Or install directly from GitHub
openclaw plugins install github:Morpheis/openclaw-recall

# Or clone and install locally
git clone https://github.com/Morpheis/openclaw-recall.git
openclaw plugins install ./openclaw-recall

Manual install

# Copy to extensions directory
mkdir -p ~/.openclaw/extensions/openclaw-recall
cp -r /path/to/openclaw-recall/* ~/.openclaw/extensions/openclaw-recall/

# Add to your openclaw.json

{
  "plugins": {
    "allow": ["openclaw-recall"],
    "entries": {
      "openclaw-recall": {
        "enabled": true
      }
    }
  }
}

Then restart:

openclaw gateway restart

Verify it's working

# Check plugin loaded
openclaw plugins list | grep recall

# Watch recall in real-time
openclaw logs --follow | grep recall

# Or use the /recall command in chat
/recall

Requirements

• OpenClaw 2026.3.x or later
• An embedding provider for memory search (Voyage, OpenAI, Gemini, or Mistral API key configured in OpenClaw)
• sqlite3 CLI for LCM source (pre-installed on macOS/Linux)
• qmd CLI for QMD source (optional: npm install -g @tobilu/qmd)

How It Works

Message arrives → before_prompt_build hook fires
                      │
                ┌─────▼──────────┐
                │ Skip Checks    │  Heartbeats, short msgs, casual msgs
                └─────┬──────────┘
                      │
                ┌─────▼──────────┐
                │ Extract Query  │  Proper nouns + adaptive strategy
                └─────┬──────────┘
                      │
        ┌─────────────┴─────────────┐
        │ (parallel)                │ (parallel)
        ▼                           ▼
  ┌───────────────┐          ┌───────────┐
  │ Text Recall   │          │  Engram   │
  │ ┌────┐┌────┐  │          │ Knowledge │
  │ │mem ││LCM │  │          │  Graph    │
  │ └────┘└────┘  │          │  Search   │
  │ ┌────┐        │          └─────┬─────┘
  │ │QMD │        │                │
  │ └────┘        │                │
  │  ↓ rank/dedup │                │
  │  ↓ filter     │                │
  │  ↓ top N      │                │
  └───────┬───────┘                │
          │                        │
          ▼                        ▼
  ## 🧠 Recalled Context   ## 🔗 Knowledge Graph
  (text snippets)           (nodes + edges)
          │                        │
          └────────┬───────────────┘
                   ▼
          { prependContext } → injected into model context
          (sections omitted individually if no results)

The model sees a block like this at the top of its context:

## 🧠 Recalled Context
Automatically retrieved based on the user's message.
Use for context; search for more detail if needed.

- **[memory/2026-03-15.md:42-48]** Discussed monorepo PR, Gemini review
  suggested printf over echo for cross-platform compatibility
- **[lcm:messages/2525]** [assistant] Weekly summary written...

## 🔗 Knowledge Graph
- **[engram:acme/AuthService]** AuthService (service) → calls: UserDB, TokenCache ← called_by: API-Gateway
- **[engram:acme/API-Gateway]** API-Gateway (service) → calls: AuthService, OrderService ← called_by: Frontend

Each section is independent — if text recall finds nothing but engrams match, only the Knowledge Graph section appears (and vice versa). Nothing is injected when both are empty — zero overhead.

How much context does it use?

Recall is controlled by a single setting: maxSnippets (default: 5). Each snippet is truncated to ~250 characters at a sentence boundary. With 5 snippets, recall typically adds ~300-400 tokens to the context — a fraction of most context windows.

There's no percentage-based token budget. You control exactly how many results you want, and the plugin takes the best ones.

Memory Sources

Recall searches multiple systems in parallel and merges the results. Enable what you have.

Memory Search (built-in)

Searches OpenClaw's built-in memory index via openclaw memory search. Works with any configured embedding provider (Voyage, OpenAI, etc.).

This is the universal source — if you only enable one, enable this one.

"sources": {
  "memorySearch": {
    "enabled": true,
    "maxResults": 5
  }
}

Note: If you use the memory-lancedb plugin (which has its own auto-recall), the memory-search source is automatically disabled to avoid duplicate work. LCM and QMD sources still run since lancedb doesn't cover them. Override this with "skipMemorySearchIfLancedb": false.

LCM (conversation history)

Searches the Lossless Context Management database via SQLite FTS5 full-text search. Finds relevant past conversations — including messages that have been compacted out of the active context window.

Requires the lossless-claw plugin.

"sources": {
  "lcm": {
    "enabled": true,
    "scope": "messages",
    "maxResults": 3
  }
}

Scope options:

• "messages" — search raw conversation messages (default, most useful)
• "summaries" — search compacted summary nodes
• "both" — search everything

QMD (local search index)

Searches a QMD index via BM25 (fast keyword search, no API calls). Useful for agents with large file collections already indexed by QMD.

"sources": {
  "qmd": {
    "enabled": true,
    "command": "qmd search",
    "maxResults": 5
  }
}

Engram (knowledge graphs) — Dedicated Section

Searches Engram knowledge graphs — persistent relationship graphs of codebases, organizations, infrastructure, and concepts.

Unlike other sources, engram results get their own ## 🔗 Knowledge Graph section rather than competing in the ranked recall pool. Graph context is structural (nodes + edges) — fundamentally different from text snippets — so it shouldn't compete on relevance score. Engram search runs in parallel with text recall, adding no extra latency.

If no engram results match, the section is omitted entirely (zero extra tokens).

Example output:

## 🧠 Recalled Context
Automatically retrieved based on the user's message.

- **[memory/2026-03-25.md]** Mental Model tool built...
- **[lcm:messages/436]** Discussed API key rotation policy for staging environment...

## 🔗 Knowledge Graph
- **[engram:acme/AuthService]** AuthService (service) → calls: UserDB, TokenCache ← called_by: API-Gateway

Auto-detection: If ~/.config/engram/models.db exists, this source can auto-activate. Set enabled: true to force it on, or enabled: false to disable.

"sources": {
  "engram": {
    "enabled": true,
    "maxResults": 3,
    "excludeModels": ["personal-contacts"]
  }
}

Config options:

• enabled — true/false (default: false, auto-detects if DB exists)
• command — engram CLI command (default: engram; customize if engram isn't on your PATH)
• maxResults — max engram snippets per message (default: 3, independent of text recall's maxSnippets)
• excludeModels — array of model names to skip (e.g., personal/family models you don't want in recall)

Enabling and Disabling

Disable the entire plugin

"plugins": {
  "entries": {
    "openclaw-recall": {
      "enabled": false
    }
  }
}

Disable individual sources

Keep the plugin active but turn off specific backends:

"config": {
  "sources": {
    "memorySearch": { "enabled": true },
    "lcm": { "enabled": false },
    "qmd": { "enabled": false }
  }
}

memory-lancedb coexistence

When memory-lancedb is active, Recall automatically disables only the memory-search source (since lancedb handles its own recall for that layer). Other sources like LCM and QMD continue running — lancedb doesn't search conversation history or local file indexes.

To force memory-search on even with lancedb:

"config": {
  "skipMemorySearchIfLancedb": false
}

Exclusion Patterns

This is the most important thing to tune. Without good exclusion patterns, recall will surface noise — changelogs, license files, vendored dependencies, backup directories — instead of the context that actually matters.

Why exclusions matter

Every noisy result displaces a useful one. With a limited snippet count (default 5), one bad result means 20% of your recall budget is wasted. Invest time tuning exclusions early — watch the ## 🧠 Recalled Context block in your agent's responses for a few days and add patterns for anything irrelevant that keeps showing up. The quality improvement compounds over time.

How patterns work

There are two types of exclusions:

excludeFiles — exact filename matches (case-insensitive). These are files already injected into your agent's context by OpenClaw's bootstrap system. No point recalling what's already there.

"excludeFiles": [
  "AGENTS.md", "SOUL.md", "TOOLS.md", "IDENTITY.md",
  "USER.md", "HEARTBEAT.md", "MEMORY.md", "PERSONALITY.md"
]

excludePatterns — glob-style path patterns with * and ** wildcards (case-insensitive). These filter out entire categories of files. Custom patterns are additive — built-in defaults are always included, so you only need to add patterns specific to your setup.

"excludePatterns": [
  "**/my-backup/**",
  "**/*-venv/**",
  "**/pipeline/data/**",
  "**/zh-cn/**"
]

Built-in defaults

These patterns are always active (your custom patterns add to them, never replace them):

**/node_modules/**     # npm/yarn dependencies
**/dist/**             # compiled output
**/build/**            # build artifacts
**/lib/**              # vendored libraries
**/vendor/**           # vendored dependencies
**/e2e/**              # end-to-end suites are high-noise for auto-recall
**/prompts/**          # prompt libraries crowd out task-relevant context
**/*transcript*.md     # long transcripts rarely help auto-recall ranking
lossless-claw/**       # keep lossless-claw docs for explicit/manual search, not auto-recall
**/changelog.md        # release notes (noisy keyword overlap)
**/CHANGELOG.md
**/license.md          # license text
**/LICENSE.md
**/contributing.md     # contribution guides
**/CONTRIBUTING.md
**/code-of-conduct.md  # codes of conduct
**/CODE_OF_CONDUCT.md
**/security.md         # security policies
**/SECURITY.md

Pattern syntax

Tips for tuning exclusions

1. Start broad, refine later. Begin with the defaults and add patterns as you notice noise in the recalled context block.
2. Watch the recalled context block. If you see the same irrelevant source repeatedly, add a pattern for it.
3. Exclude backup directories. Backup mirrors cause duplicate results that displace unique content.
4. Exclude generated files. Build output, compiled code, and auto-generated content rarely contains useful recall context.
5. Exclude localization you don't use. i18n directories in other languages add noise to keyword searches.
6. Exclude large data directories. Raw datasets, logs, and pipeline data overwhelm useful results.
7. Be careful with blanket docs/ exclusions. Docs are often useful, especially for "how does this work?" questions. Prefer pruning clearly noisy subtrees first.
8. Prefer diversity over doc spam. Recall caps docs/specs/prompts to one snippet per repo subtree (for example openclaw/docs/**) so a single documentation tree doesn't monopolize the snippet budget.

Full Configuration Reference

All settings go under plugins.entries.openclaw-recall.config:

{
  "plugins": {
    "entries": {
      "openclaw-recall": {
        "enabled": true,
        "config": {
          // Max snippets injected per message (default: 5)
          // This is the primary knob for controlling recall volume.
          // Lower = tighter, higher quality. Higher = more context surfaced.
          "maxSnippets": 5,

          // Minimum relevance score 0-1 to include a result (default: 0.3)
          // Raise this if you're getting too much noise
          "minRelevanceScore": 0.3,

          // Skip recall for very short messages after metadata stripping (default: enabled, 40 chars)
          // Proper nouns in short messages override this (e.g., "OathKeeper?" still triggers)
          "skipShortMessages": {
            "enabled": true,
            "minChars": 40
          },

          // Skip heartbeat messages entirely (default: true)
          "skipHeartbeats": true,

          // Only skip memory-search when lancedb is active (default: true)
          // Other sources (LCM, QMD) still run since lancedb doesn't cover them
          "skipMemorySearchIfLancedb": true,

          // Files already in context — excluded from results
          "excludeFiles": [
            "AGENTS.md", "SOUL.md", "TOOLS.md", "IDENTITY.md",
            "USER.md", "HEARTBEAT.md", "MEMORY.md", "PERSONALITY.md"
          ],

          // Glob patterns to exclude (additive — built-in defaults always included)
          "excludePatterns": [
            "**/my-backup/**",
            "**/*-venv/**"
          ],

          // Memory sources — enable what you have
          "sources": {
            "memorySearch": { "enabled": true, "maxResults": 5 },
            "lcm": { "enabled": true, "scope": "messages", "maxResults": 3 },
            "qmd": { "enabled": false, "command": "qmd search", "maxResults": 5 },
            "engram": {
              "enabled": true,       // auto-detects if DB exists when omitted
              "maxResults": 3,
              "excludeModels": [],   // model names to skip (e.g., "personal-contacts")
              "command": "engram"
            }
          },

          // Debug logging (writes NDJSON to workspace)
          "debug": {
            "enabled": false,
            "logPath": ".recall/debug.log"
          }
        }
      }
    }
  }
}

How Recall Differs from memory-lancedb

Both are valid approaches. Recall is designed for agents that already have memory systems (OpenClaw memory, LCM, QMD) and want automatic retrieval without adding another database.

Troubleshooting

Plugin not loading

• Check plugins.allow includes "openclaw-recall"
• Check plugins.entries.openclaw-recall.enabled is true
• Run openclaw plugins list | grep recall — should show "loaded"
• Run openclaw doctor for compatibility warnings

No results appearing

• Check your embedding provider is configured: openclaw memory status
• Try a manual search: openclaw memory search --json --query "test"
• Heartbeat messages are skipped by default
• Short messages (<40 chars after metadata stripping) are skipped unless they contain proper nouns
• If all sources are disabled, nothing will be searched

Poor or noisy results

• Most common fix: add exclusion patterns for noisy directories
• Raise minRelevanceScore (try 0.4 or 0.5)
• Lower maxSnippets to keep only the highest-quality results
• Enable debug logging to see what's being searched and returned

Duplicate results

• Add backup/mirror directories to excludePatterns
• Check QMD collections for overlapping indexed paths
• Cross-source dedup normalizes paths but can't catch all mirrors

Commands

Roadmap

• Memory search source (OpenClaw built-in)
• LCM source (conversation history via FTS5)
• QMD source (local BM25 search)
• Proper noun prioritization in query extraction
• Metadata envelope stripping
• Cross-collection deduplication with path normalization
• Short message / casual message filtering
• LCM tool call/result filtering (noise reduction)
• LCM content cleaning (strip metadata from results)
• memory-lancedb coexistence (skip only memory-search, not entire plugin)
• Explicit capability registration (before_prompt_build + /recall command)
• Engram source (knowledge graph search with auto-detection)
• npm publish / ClawHub listing
• Recall quality analytics (track hits/misses across sessions)
• Per-channel config (different settings per channel)

License

MIT