Release March 3, 2026

OpenClaw v2026.3.2: What's New and How to Upgrade

The latest OpenClaw release ships improved memory management, faster vector lookups, and a new tool execution sandbox. Here's everything you need to know.

AgentHost Team

OpenClaw v2026.3.2 is now available and natively supported on AgentHost. This release focuses on memory reliability, inference speed, and execution safety — three areas that matter most for production agent deployments. Here’s a detailed breakdown of what changed and how to migrate.

Memory Management Overhaul

The most impactful change in v2026.3.2 is a complete rewrite of the context window manager. Previous versions used a naive truncation strategy when approaching token limits: oldest messages were dropped first. This worked for short tasks but caused catastrophic failures in long-running agents that needed to reference early context for later decisions.

The new Hierarchical Context Compressor (HCC) takes a different approach. Rather than truncating, it summarizes older context blocks using a lightweight 7B model running locally, preserving semantic content while reducing token count by up to 70%. The compressed summary is injected back into the context window as a synthetic message, maintaining continuity without the full token cost.

To enable HCC in your configuration:

memory:
  strategy: hierarchical
  compression_model: openclaw-compress-7b
  compression_threshold: 0.75  # compress when context is 75% full
  summary_max_tokens: 512

Faster Vector Lookups

Long-term memory retrieval via vector search has been a bottleneck in previous versions. v2026.3.2 ships with a new HNSW index implementation that reduces p99 retrieval latency by approximately 40% on collections larger than 100k embeddings.

The new index is opt-in for existing deployments to avoid reindexing costs:

vector_store:
  index_type: hnsw  # previously: flat
  hnsw_ef_construction: 200
  hnsw_m: 16

New deployments use HNSW by default. Existing flat indexes continue to work but will not receive future performance improvements.

Tool Execution Sandbox

v2026.3.2 introduces a sandboxed tool execution environment for code interpreter and shell tool calls. Previously, tool execution happened in the same process space as the agent runtime, meaning a runaway subprocess could affect agent stability.

The new sandbox runs each tool call in an isolated cgroup with:

This is enabled by default for all new deployments. To configure limits:

tool_sandbox:
  enabled: true
  memory_limit_mb: 512
  cpu_timeout_seconds: 30
  network_egress: false
  allowed_paths:
    - /workspace
    - /tmp

Breaking Changes

Python SDK: The Agent.run_sync() method has been removed. Use asyncio.run(Agent.run()) instead.

Configuration: The context.max_tokens field is deprecated in favor of memory.compression_threshold. The old field still works but logs a deprecation warning.

Tool responses: Tool results are now wrapped in a ToolResult object with content, metadata, and execution_time_ms fields. Code that directly accesses tool_response as a string will need to be updated to tool_response.content.

Upgrading on AgentHost

AgentHost deployments on v2026.3.1 or earlier will be automatically offered the upgrade via the dashboard. The upgrade process:

  1. Snapshot your current deployment state
  2. Pull the v2026.3.2 base image
  3. Run configuration migration with openclaw migrate --from 2026.3.1
  4. Verify with a dry-run before promoting to production

The migration script handles the max_tokens → compression_threshold conversion automatically. Review the output before committing.

Zero-downtime upgrades are supported via blue/green deployment — spin up a v2026.3.2 instance alongside your current deployment, validate behavior, then switch traffic.

All Posts