palaia 2.0.4

Local, cloud-free memory for OpenClaw agents.

Palaia — The Knowledge OS for OpenClaw Agent Teams

Crash-safe. Local-first. Zero-cloud. The memory system that makes your agents smarter over time.

---

Install — Zero Config

pip install "palaia[fastembed]"

OpenClaw plugin (3 lines in your config):

{
  "plugins": {
    "load": { "paths": ["<npm-global>/node_modules/@byte5ai/palaia"] },
    "allow": ["palaia"],
    "slots": { "memory": "palaia" }
  }
}

That's it. Palaia works immediately — memories are injected into every prompt, significant exchanges are captured automatically, and semantic recall finds what's relevant. No config edits needed.

Upgrading from v1.x? Run palaia doctor --fix to get the new optimized defaults.

Quick Start

palaia init                                         # Initialize store
palaia write "API rate limit is 100 req/min" \
  --type memory --tags api,limits                   # Save knowledge
palaia query "what's the rate limit"                # Find it by meaning

---

Why Palaia?

WAL-Backed Crash Safety

Every write goes through a write-ahead log before touching storage. Power loss mid-write? Palaia recovers automatically on next startup. No data loss. No corruption. No "oops."

Intelligent Tiering

Memories automatically organize by usage: HOT (active), WARM (recent), COLD (archived). Frequently accessed entries stay fast. Nothing gets deleted — old memories fade to background storage.

Structured Entry Types

Not all knowledge is equal. Classify entries as memory (facts, decisions), process (workflows, checklists), or task (action items with status, priority, assignee). Query by type for focused results.

Multi-Agent Collaboration

Multiple agents share one store with scope-based access control. Private entries stay private. Team entries are shared. Projects group related knowledge. Inter-agent memos enable async communication.

Zero-Cloud Architecture

Everything runs on your machine. No API keys required for core functionality. No database server. No cloud dependency. Your data never leaves your infrastructure.

Adaptive Nudging

Palaia teaches agents good habits through CLI output hints — then stops once they learn. The graduation system tracks consecutive successes and retires nudges when agents demonstrate independence. Regression detection re-activates nudges if habits slip.

OpenClaw-Native

Built as a first-class OpenClaw plugin. Auto-capture of significant exchanges. Query-based contextual recall before each prompt. LLM-powered knowledge extraction. Configurable capture levels from conservative to aggressive.

---

Features

Feature	Details
Semantic Search	Find by meaning, not keywords. Providers: fastembed, sentence-transformers, OpenAI, Gemini, Ollama, BM25
Crash-Safe Writes	WAL-backed — survives power loss, kills, OOM
Auto-Capture	OpenClaw plugin captures significant exchanges automatically
Structured Types	memory, process, task — with status, priority, assignee fields
Multi-Agent	Shared store, scopes (private/team/public), agent aliases, inter-agent memos
Smart Tiering	HOT → WARM → COLD rotation based on access patterns
Garbage Collection	Automatic tier rotation, WAL cleanup, stale entry management
OpenClaw Plugin	Drop-in replacement for built-in memory — query-based recall, auto-capture, LLM extraction
Projects	Group entries by project with default scopes and ownership
Document Ingestion	Index PDFs, HTML, Markdown for RAG search
Adaptive Nudging	Teaches agents best practices, graduates when they learn

---

Comparison

Feature	Palaia	Stock Memory	Mem0	Engram
Local-first	Yes	Yes	No (cloud)	Yes
Crash-safe (WAL)	Yes	No	N/A	No
Auto-Capture	Yes (plugin)	No	Yes	No
Structured Types	Yes (memory/process/task)	No	No	No
Multi-Agent Scopes	Yes (private/team/public)	No	Per-user	No
Smart Tiering	Yes (HOT/WARM/COLD)	No	No	No
Garbage Collection	Yes (automatic)	Manual	Managed	Manual
OpenClaw Plugin	Native	Built-in	No	No
Semantic Search	Hybrid (embedding + BM25)	None	Embedding	Embedding
Zero-Cloud	Yes	Yes	No	Yes

---

Configuration

Plugin Config (OpenClaw)

Set in openclaw.json under plugins.entries.palaia.config:

Key	Default	Description
memoryInject	true	Inject relevant memories into agent context
maxInjectedChars	8000	Max characters for injected memory context
autoCapture	true	Capture significant exchanges automatically
captureFrequency	"significant"	"every" or "significant"
captureMinTurns	2	Minimum turns before capture
captureModel	auto	Model for LLM extraction (e.g. "anthropic/claude-haiku-3")
recallMode	"query"	"list" (tier-based) or "query" (semantic)
recallTypeWeight	{process:1.5, task:1.2, memory:1.0}	Type-aware result weighting

Capture Levels

Configure via palaia init --capture-level:

Level	autoCapture	Frequency	Min Turns
off	false	—	—
sparsam	true	significant	5
normal	true	significant	2
aggressiv	true	every	1

---

CLI Reference

palaia init [--agent NAME] [--capture-level LEVEL]   Initialize store
palaia write "text" [--type TYPE] [--tags a,b]       Save a memory
palaia query "search" [--type TYPE] [--project P]    Search by meaning
palaia get <id>                                       Read specific entry
palaia list [--tier T] [--type T] [--status S]       List entries
palaia edit <id> [--status done]                      Edit entry
palaia status                                         System health
palaia doctor [--fix]                                 Diagnose + fix
palaia project create|list|show|query|delete          Manage projects
palaia memo send|inbox|ack|broadcast                  Inter-agent messaging
palaia ingest <source> [--project P]                  Index documents (RAG)
palaia detect                                         Available providers
palaia warmup                                         Pre-build search index
palaia migrate [--suggest]                            Import / suggest types

All commands support --json for machine-readable output.

---

Development

git clone https://github.com/iret77/palaia.git
cd palaia
pip install -e ".[dev]"
pytest

640+ tests. Contributions welcome.

Links

• GitHub — Source + Issues
• PyPI — Package registry
• ClawHub — Install via agent skill
• OpenClaw — The agent platform Palaia is built for
• CHANGELOG — Release history

---

MIT — (c) 2026 byte5 GmbH