Jasper Recall v0.2.3

Local RAG (Retrieval-Augmented Generation) system for AI agent memory. Gives your agent the ability to remember and search past conversations.

New in v0.2.2: Shared ChromaDB Collections — separate collections for private, shared, and learnings content. Better isolation for multi-agent setups.

New in v0.2.1: Recall Server — HTTP API for Docker-isolated agents that can't run CLI directly.

New in v0.2.0: Shared Agent Memory — bidirectional learning between main and sandboxed agents with privacy controls.

When to Use

• - Memory recall: Search past sessions for context before answering
• Continuous learning: Index daily notes and decisions for future reference
• Session continuity: Remember what happened across restarts
• Knowledge base: Build searchable documentation from your agent's experience

Quick Start

Setup

One command installs everything:

CODEBLOCK0

This creates:

• - Python venv at INLINECODE0
• ChromaDB database at INLINECODE1
• CLI scripts in INLINECODE2
• OpenClaw plugin config in INLINECODE3

Why Python?

The core search and embedding functionality uses Python libraries:

• - ChromaDB — Vector database for semantic search
• sentence-transformers — Local embedding models (no API needed)

These are the gold standard for local RAG. There are no good Node.js equivalents that work fully offline.

Why a Separate Venv?

The venv at ~/.openclaw/rag-env provides:

The dependencies are heavy (~200MB total with the embedding model), but this is a one-time download that runs entirely locally.

Basic Usage

Search your memory:
CODEBLOCK1

Index your files:
CODEBLOCK2

Create session digests:
CODEBLOCK3

How It Works

Three Components

1. 1. digest-sessions — Extracts key info from session logs (topics, tools used)
2. index-digests — Chunks and embeds markdown files into ChromaDB
3. recall — Semantic search across your indexed memory

What Gets Indexed

By default, indexes files from ~/.openclaw/workspace/memory/:

• - *.md — Daily notes, MEMORY.md
• INLINECODE7 — Session summaries
• INLINECODE8 — Project documentation
• INLINECODE9 — Development logs (if present)

Embedding Model

Uses sentence-transformers/all-MiniLM-L6-v2:

• - 384-dimensional embeddings
• ~80MB download on first run
• Runs locally, no API needed

Agent Integration

Memory-Augmented Responses

CODEBLOCK4

Automated Indexing (Heartbeat)

Add to HEARTBEAT.md:
CODEBLOCK5

Cron Job

Schedule regular indexing:
CODEBLOCK6

Shared Agent Memory (v0.2.0+)

For multi-agent setups where sandboxed agents need access to some memories:

Memory Tagging

Tag entries in daily notes:

CODEBLOCK7

ChromaDB Collections (v0.2.2+)

Memory is stored in separate collections for isolation:

Collection selection:
CODEBLOCK8

Sandboxed Agent Access

CODEBLOCK9

Moltbook Agent Setup (v0.4.0+)

For the moltbook-scanner (or any sandboxed agent), use the built-in setup:

CODEBLOCK10

This creates:

• - ~/bin/recall — Wrapper that forces --public-only flag
• INLINECODE17 — Symlink to main workspace's shared memory

The sandboxed agent can then use:
CODEBLOCK11

Privacy model:

1. 1. Main agent tags memories as [public] or [private] in daily notes
2. INLINECODE20 extracts [public] content to INLINECODE22
3. Sandboxed agents can ONLY search the shared collection

Privacy Workflow

CODEBLOCK12

CLI Reference

recall

CODEBLOCK13

serve (v0.2.1+)

CODEBLOCK14

Example (from Docker container):
CODEBLOCK15

privacy-check (v0.2.0+)

CODEBLOCK16

sync-shared (v0.2.0+)

CODEBLOCK17

index-digests

CODEBLOCK18

digest-sessions

CODEBLOCK19

Configuration

Custom Paths

Set environment variables:

CODEBLOCK20

Chunking

Default settings in index-digests:

• - Chunk size: 500 characters
• Overlap: 100 characters

Security Considerations

⚠️ Review these settings before enabling in production:

Server Binding

The serve command defaults to 127.0.0.1 (localhost only). Do not use --host 0.0.0.0 unless you explicitly intend to expose the API externally and have secured it appropriately.

Private Memory Access

The server enforces public_only=true by default. The env var RECALL_ALLOW_PRIVATE=true bypasses this restriction. Never set this on public/shared hosts — it exposes your private memories to any client.

autoRecall Plugin

When autoRecall: true in the OpenClaw plugin config, memories are automatically injected before every agent message. Consider:

• - Set publicOnly: true in plugin config for sandboxed agents
• Review which collections will be searched
• Use minScore to filter low-relevance injections

What's automatically skipped (no recall triggered):

• - Heartbeat polls (HEARTBEAT, Read HEARTBEAT.md, HEARTBEAT_OK)
• Messages containing INLINECODE35
• Messages < 10 characters
• Agent-to-agent messages (cron jobs, workers, spawned agents)
• Automated reports (📋 PR Review, 🤖 Codex Watch, ANNOUNCE_*)
• Messages from senders starting with agent: or INLINECODE40

Safer config for untrusted contexts:
CODEBLOCK21

Environment Variables

The following env vars affect behavior — set them explicitly rather than relying on defaults:

Dry-Run First

Before sharing or syncing, use dry-run options to preview what will be exposed:

CODEBLOCK22

Sandboxed Environments

For maximum isolation, run jasper-recall in a container or dedicated account:

• - Limits risk of accidental data exposure
• Separates private memory from shared contexts
• Recommended for multi-agent setups with untrusted agents

Troubleshooting

"No index found"
CODEBLOCK23

"Collection not found"
CODEBLOCK24

Model download slow
First run downloads ~80MB model. Subsequent runs are instant.

Links

• - GitHub: https://github.com/E-x-O-Entertainment-Studios-Inc/jasper-recall
• npm: https://www.npmjs.com/package/jasper-recall
• ClawHub: https://clawhub.ai/skills/jasper-recall