palaia

Knowledge System for
Agent Teams

Shared, persistent memory — across sessions, projects, and scopes.

OpenClaw · Claude Code — local, crash-safe, open source.

Why palaia

What you can do with palaia.

Your AI agents are powerful — but they forget everything between sessions. palaia changes that.

Your agents learn and remember

Every conversation builds knowledge. Your agents remember decisions, preferences, and context — across sessions, across projects. You stop repeating yourself. They start getting smarter.

Your agent team works as a team

One agent discovers a bug fix. The others know about it immediately. No duplicate work, no conflicting changes, no lost context. Your agents share knowledge like a real team — with structure, not chaos.

Your data never leaves your machine

No cloud. No external APIs for storage. Everything stays on your filesystem. Your code, your client data, your business logic — under your control, always. No surprise bills, no vendor lock-in.

Works with your tools

Claude Desktop, Cursor, OpenClaw — or any MCP-capable tool. One pip install and a config line is all it takes. No account needed, no Docker, no database to set up. Under 2 minutes from zero to persistent memory.

Your agents know what matters

Not all memories are equal. palaia lets each agent control which knowledge gets injected — per agent, per project. High-noise entries from one agent don't pollute another's context.

Migration without the mess

Hundreds of accumulated entries? palaia curate clusters them thematically, finds duplicates, and lets you decide per cluster — not per entry. Edit a Markdown report, apply it, done.

Gets faster, not slower

Native vector search via sqlite-vec — not Python cosine similarity. Background embed server keeps the model in RAM. Sub-second queries even with thousands of entries.

The difference

From memory to knowledge.

Team coordination
Generic memory

Agent A writes a fix. Agent B doesn't know and works on the same issue. Duplicate work, merge conflicts, wasted tokens.

With palaia

Project locking prevents parallel work on the same repo. Team-scoped entries visible to all agents. Inter-agent memos for direct coordination.

Knowledge structure
Generic memory

Everything stored as flat text. "Customer prefers CSV" lives next to "deploy runs on port 3000" — no types, no priority, no status tracking.

With palaia

Memory, Process, Task — typed entries with status, priority, and assignee. Your agent knows the difference between a fact and an open task.

Agent discipline
Generic memory

Your agent needs constant reminders to save context. You prompt it every time: "remember this", "write that down". Manual, fragile, forgotten.

With palaia

Adaptive nudging trains your agents to self-document. Reminders fade as behavior is learned. The longer you use it, the smarter it gets.

Context control
Generic memory

Agent A's noisy debug logs flood Agent B's carefully curated project knowledge. No way to filter.

With palaia

Each agent has injection priorities. Block entries, weight types, set score thresholds — per agent, per project. The orchestrator sees strategy docs, the coder sees API specs.

Why palaia

What sets palaia apart.

Multi-Agent Coordination

Inter-agent memos, project locking to prevent parallel work on the same repo, and scope isolation (private / team / public). Built for teams, not solo agents.

Zero-Cloud

Everything local. Knowledge stored as plain text files on your filesystem — SQLite holds the search index and vectors. No API keys, no data ever leaves your machine. Zero-config after install.

Crash-Safe (WAL)

Write-Ahead Logging like a production database. No data loss on crashes or restarts — production-grade crash safety that few local memory systems can match.

Structured Knowledge Types

Memory, Process, Task — with status, priority, and assignee. Decisions and workflows are first-class citizens, not text blobs.

Session Continuity

Auto-briefing loads your last session summary on startup. Auto-summaries are saved at session end via LLM or rule-based capture. Your agent always knows where you left off — no manual context-setting.

Adaptive Nudging

Trains your agents to self-document. Nudges are shown until the behavior is learned — then they stop. The longer you use palaia, the more efficient it gets.

Injection Priorities

Per-agent, per-project control over what gets injected. Block entries, adjust type weights, set minimum scores. Your orchestrator doesn't need your debugger's logs.

Automatic Tiering

HOT / WARM / COLD — frequently accessed knowledge stays fast, old entries are archived, never deleted. Active context always surfaces first.

Knowledge Curation

Cluster, deduplicate, and clean accumulated knowledge. Markdown report as interface — edit 20 cluster decisions instead of 500 individual entries. Export as portable package.

Native Vector Search

sqlite-vec for local SIMD-accelerated search, pgvector for PostgreSQL teams. ~30x faster than Python cosine similarity on large stores. Scales with your knowledge, not against it.

Sub-Second Queries

Background embed server keeps the model in RAM. No more 5-second cold starts per query. Your agent gets answers in under 500ms — memory access that doesn't break the flow.

Document Ingestion

Ingest files, URLs, and directories. Automatic chunking and embedding, stored as regular entries. PDF support with source attribution per chunk.

Memory Source Footnotes

Agent responses show where knowledge came from — your user sees which memories influenced the answer. Builds trust and transparency.

OpenClaw Plugin

Deep lifecycle integration via ContextEngine: 7 hooks from bootstrap to subagent management. Auto-capture after every turn, auto-recall before every prompt.

MCP Server

Standalone memory server for Claude Desktop, Cursor, and any MCP-capable tool. 7 tools: search, store, read, edit, list, status, gc. Read-only mode for untrusted hosts. pip install and one config line.

ContextEngine Integration

Seven lifecycle hooks for deep OpenClaw integration: bootstrap, ingest, assemble, compact, afterTurn, prepareSubagentSpawn, onSubagentEnded. Goes beyond simple prompt injection.

Not just memory. Not a cloud service. Not a RAG pipeline.
palaia is the knowledge system for agent teams — crash-safe, local, instantly ready.

Honest comparison

How palaia compares.

Verified facts only. No marketing claims.

supported~partial / with caveatsnot supported
palaia15/17
Local-firstMCP ServerCrash-Safe (WAL)Native Vector SearchSub-Second QueriesSQLite BackendPostgreSQL BackendStructured TypesMulti-Agent ScopesPer-Agent PrioritiesKnowledge CurationAuto-CaptureAdaptive NudgingMIT LicenseKnowledge GraphBi-temporal QueriesHybrid Search
MemPalace9/17
Local-firstMCP ServerCrash-Safe (WAL)Native Vector SearchSub-Second QueriesSQLite BackendPostgreSQL BackendStructured TypesMulti-Agent ScopesPer-Agent PrioritiesKnowledge CurationAuto-CaptureAdaptive NudgingMIT LicenseKnowledge GraphBi-temporal QueriesHybrid Search
Mem04/17
Local-firstMCP ServerCrash-Safe (WAL)Native Vector SearchSub-Second QueriesSQLite BackendPostgreSQL BackendStructured TypesMulti-Agent ScopesPer-Agent PrioritiesKnowledge CurationAuto-CaptureAdaptive NudgingMIT LicenseKnowledge GraphBi-temporal QueriesHybrid Search
Graphiti10/17
Local-firstMCP ServerCrash-Safe (WAL)Native Vector SearchSub-Second QueriesSQLite BackendPostgreSQL BackendStructured TypesMulti-Agent ScopesPer-Agent PrioritiesKnowledge CurationAuto-CaptureAdaptive NudgingMIT LicenseKnowledge GraphBi-temporal QueriesHybrid Search
Cognee10/17
Local-firstMCP ServerCrash-Safe (WAL)Native Vector SearchSub-Second QueriesSQLite BackendPostgreSQL BackendStructured TypesMulti-Agent ScopesPer-Agent PrioritiesKnowledge CurationAuto-CaptureAdaptive NudgingMIT LicenseKnowledge GraphBi-temporal QueriesHybrid Search
Letta5/17
Local-firstMCP ServerCrash-Safe (WAL)Native Vector SearchSub-Second QueriesSQLite BackendPostgreSQL BackendStructured TypesMulti-Agent ScopesPer-Agent PrioritiesKnowledge CurationAuto-CaptureAdaptive NudgingMIT LicenseKnowledge GraphBi-temporal QueriesHybrid Search
claude-mem6/17
Local-firstMCP ServerCrash-Safe (WAL)Native Vector SearchSub-Second QueriesSQLite BackendPostgreSQL BackendStructured TypesMulti-Agent ScopesPer-Agent PrioritiesKnowledge CurationAuto-CaptureAdaptive NudgingMIT LicenseKnowledge GraphBi-temporal QueriesHybrid Search
OpenClaw (native)2/17
Local-firstMCP ServerCrash-Safe (WAL)Native Vector SearchSub-Second QueriesSQLite BackendPostgreSQL BackendStructured TypesMulti-Agent ScopesPer-Agent PrioritiesKnowledge CurationAuto-CaptureAdaptive NudgingMIT LicenseKnowledge GraphBi-temporal QueriesHybrid Search
Claude (native)3/17
Local-firstMCP ServerCrash-Safe (WAL)Native Vector SearchSub-Second QueriesSQLite BackendPostgreSQL BackendStructured TypesMulti-Agent ScopesPer-Agent PrioritiesKnowledge CurationAuto-CaptureAdaptive NudgingMIT LicenseKnowledge GraphBi-temporal QueriesHybrid Search

v2.8 — verified May 2026 from GitHub repos + official docs. ~ = partial: Mem0 local-first defaults to cloud LLM, requires Ollama for full local; Mem0 WAL is transactional, not journal mode; Mem0 SQLite is history/audit only. Graphiti runs against external graph DBs (Neo4j / FalkorDB / Kuzu / Neptune) — no SQLite/PostgreSQL native. Cognee auto-capture via Claude Code plugin hooks. Letta is cloud-first with Docker self-host option (PostgreSQL primary). claude-mem vector search uses ChromaDB (separate process); multi-agent is session-isolation only. OpenClaw (native) and Claude (native) are baselines — no third-party install, minimal feature coverage. Licenses: palaia MIT, MemPalace MIT, Mem0 / Graphiti / Cognee / Letta / claude-mem Apache 2.0, OpenClaw MIT, Claude proprietary.

How it works

Running in under 2 minutes.

01

Pick your tool, install

palaia works as an MCP server for Claude Desktop, Cursor — or as an OpenClaw plugin. Choose your path.

OpenClaw Plugin

agent prompt

Install or update the Palaia memory skill from ClawHub to the latest version (even if already present). Read the SKILL.md completely and follow it step by step. Run palaia init, then palaia doctor --fix and resolve all warnings — don't stop until the doctor report is clean. Set up completely.
Claude Code

agent prompt

Install palaia for persistent memory in this Claude Code environment. Run: pip install "palaia[mcp,fastembed]" && palaia init && palaia setup claude-code --global. Then tell me to restart Claude Code so the MCP tools become active. After restart, read the CLAUDE.md and follow its instructions.
MCP ServerClaude Desktop · Cursor · other MCP hosts
$ pip install "palaia[mcp]"

Add to your MCP config:

{
  "mcpServers": {
    "palaia": {
      "command": "palaia-mcp"
    }
  }
}

These tools require manual agent instruction — palaia provides the memory tools, but the agent doesn't learn proactive usage automatically.

02

Your agent writes knowledge

$ palaia write --type process "Deploy: merge to main triggers build + deploy to prod" --tags "ops"

Not just text — typed entries (memory, process, task) with tags, scopes, and automatic tiering. Crash-safe via WAL.

# Knowledge curation for migration
$ palaia curate analyze --output report.md
# Edit report.md, then:
$ palaia curate apply report.md
03

Your agent recalls and coordinates

$ palaia query "deployment process" --scope team

Native vector search across your entire knowledge base. Sub-second results. Project locking prevents conflicts. Memos enable agent-to-agent communication.

# Control what each agent sees
$ palaia priorities block <entry-id> --agent coder
$ palaia priorities set recall_min_score 0.6 --agent orchestrator
alt

Or install standalone

$ pip install "palaia[fastembed]" && palaia init && palaia doctor --fix

Works without MCP or OpenClaw too. Full CLI access, same features.

Under the hood

Built for reliability.

StorageText files (knowledge) + SQLite index/vectors, PostgreSQL (optional)
Searchsqlite-vec · pgvector · fastembed · BM25
PerformanceSub-second via embed server
LanguagePython 3.9+ (22,000+ LOC)
MCP Tools7 (search, store, read, edit, list, status, gc)
Tests1300+ (Python + TypeScript, CI 3.9-3.12)
IntegrationsMCP Server + OpenClaw Plugin
Version2.8

Give your agent team a shared memory.

Fully autonomous on OpenClaw and Claude Code. MCP-compatible with any other tool.

Fully autonomous — paste and go

Paste this into your OpenClaw agent
Install or update the Palaia memory skill from ClawHub to the latest version (even if already present). Read the SKILL.md completely and follow it step by step. Run palaia init, then palaia doctor --fix and resolve all warnings — don't stop until the doctor report is clean. Set up completely.

or standalone: pip install "palaia[fastembed]"see docs

View on GitHub