Назад към всички

jasper-recall

// Semantic search over indexed memory using ChromaDB. Automatically injects relevant context before agent processing.

$ git log --oneline --stat
stars:1,933
forks:367
updated:March 4, 2026
SKILL.mdreadonly

Jasper Recall - OpenClaw Plugin

Semantic search over indexed memory using ChromaDB. Automatically injects relevant context before agent processing.

Features

  • recall tool — Manual semantic search over memory
  • /recall command — Quick lookups from chat
  • /index command — Re-index memory files
  • Auto-recall — Automatically inject relevant memories before processing

Auto-Recall (The Magic ✨)

When autoRecall is enabled, jasper-recall hooks into the agent lifecycle and automatically searches your memory before every message is processed.

How It Works

┌─────────────────────────────────────────────────────────────┐
│  1. Message arrives from user                               │
│  2. before_agent_start hook fires                           │
│  3. jasper-recall searches ChromaDB with message as query   │
│  4. Results filtered by minScore (default: 30%)             │
│  5. Relevant memories injected via prependContext           │
│  6. Agent sees memories + original message                  │
│  7. Agent responds with full context                        │
└─────────────────────────────────────────────────────────────┘

What Gets Injected

<relevant-memories>
The following memories may be relevant to this conversation:
- [memory/2026-02-05.md] Worker orchestration decisions...
- [MEMORY.md] Git workflow: feature → develop → main...
- [memory/sops/codex-integration-sop.md] Codex Cloud sync...
</relevant-memories>

What's Skipped

Auto-recall won't run for:

  • Heartbeat polls (HEARTBEAT...)
  • System prompts containing NO_REPLY
  • Messages shorter than 10 characters

Configuration

In openclaw.json:

{
  "plugins": {
    "entries": {
      "jasper-recall": {
        "enabled": true,
        "config": {
          "autoRecall": true,
          "minScore": 0.3,
          "defaultLimit": 5,
          "publicOnly": false
        }
      }
    }
  }
}

Options

OptionTypeDefaultDescription
enabledbooleantrueEnable/disable plugin
autoRecallbooleanfalseAuto-inject memories before processing
minScorenumber0.3Minimum similarity score (0-1) for auto-recall
defaultLimitnumber5Default number of results
publicOnlybooleanfalseOnly search public memory (sandboxed agents)

Score Tuning

  • minScore: 0.3 — Include loosely related memories (more context, may include noise)
  • minScore: 0.5 — Only moderately relevant (balanced)
  • minScore: 0.7 — Only highly relevant (precise, may miss useful context)

Tools

recall

Manual semantic search over memory.

Parameters:

  • query (string, required): Natural language search query
  • limit (number, optional): Max results (default: 5)

Example:

recall query="what did we decide about the API design" limit=3

Returns: Formatted markdown with matching memories, scores, and sources.

Commands

/recall <query>

Quick memory search from chat.

/recall worker orchestration decisions

/index

Re-index memory files into ChromaDB. Run after updating notes.

/index

RPC Methods

For external integrations:

recall.search

{ "query": "search terms", "limit": 5 }

recall.index

Re-index memory files (no params).

Requirements

  • recall command in ~/.local/bin/
  • ChromaDB index at ~/.openclaw/chroma-db
  • Python venv at ~/.openclaw/rag-env

Installation

npx jasper-recall setup

This sets up:

  1. Python venv with ChromaDB + sentence-transformers
  2. recall, index-digests, digest-sessions scripts
  3. Initial index of memory files

When Auto-Recall Helps

✅ Great for:

  • Questions about past decisions ("what did we decide about X?")
  • Following up on previous work ("where were we with the worker setup?")
  • Context about people, preferences, projects
  • Finding SOPs and procedures

⚠️ Less useful for:

  • Brand new topics with no memory
  • Simple commands ("list files")
  • Real-time data (weather, time)

Sandboxed Agents

For agents processing untrusted input, use publicOnly:

{
  "jasper-recall": {
    "config": {
      "publicOnly": true,
      "autoRecall": true
    }
  }
}

This restricts searches to memory/shared/ and public-tagged content, preventing leakage of private memories.

Links