memory management

Error: pull model manifest: 412: The model you are attempting to pull requires a newer version of Ollama.

[Bug] Claude Opus 4.6 Model Performance Degradation

[WSL2] Ollama on Windows host unreachable from Docker container/sandbox after proxy bypass (connections timeout)

Ollama is installed on a Windows 11 host with `OLLAMA_HOST=0.0.0.0:11434`. It is healthy and reachable from the WSL2 Ubuntu 24.04 host shell. However, from inside a Docker container (NemoClaw sandbox) running inside WSL2, direct connections to the Windows-hosted Ollama instance time out after proxy bypass. This makes it impossible to use Ollama as a local inference backend from containerized workloads on the same machine.

Bug: App Router renders incorrect page and refreshes when navigating with basePath (Next.js 16)

[Bug]: If two types of api methods are used (API_KEY + localhost LLM) UI does not allow switching over.

If two types of api methods are used (API_KEY + localhost LLM) UI does not allow switching over. It will only allow the type of connection used set in primary methods. example, if it was set to use openai api key, selecting any lm-studio endpoints for API call will fail.

CPU doesn't support the bmi2 instructions

[Bug]: WhatsApp outbound media sends falsely succeed on 2026.4.5, attachment is dropped and only caption/text is delivered

On OpenClaw `2026.4.5`, outbound WhatsApp media sends can return a successful tool result with a real-looking WhatsApp `messageId`, but the recipient only receives the caption/text. The actual attachment never arrives. This affects direct `message` tool sends on WhatsApp.

ollama launch openclaw --model You still need to download the model the first time you start it when choosing a model

[Installation]: Build vllm from source fail

/api/generate returns HTTP 500 {"error":"EOF"} with qwen3.5:9b when prompt requests <tool_call> XML output

`/api/generate` works for plain text prompts and simple XML-like prompts, but can return HTTP 500 with `{"error":"EOF"}` when the prompt asks the model to emit `<tool_call>...</tool_call>` style XML output. This reproduces even without using native Ollama tools. The prompt only contains tool-like XML tags as plain text instructions. Possibly related to other Qwen XML/tool-call parsing issues, but this reproduces on `qwen3.5:9b` with `/api/generate` and without native tools.

Bug: OpenClaw 2026.3.13 npm package missing dist/gateway.js - Gateway fails to start, ERR_CONNECTION_REFUSED

Hi OpenClaw team, I'm on Windows 10/11, Node v20.20.1 (also tried v22.9.0), npm install -g openclaw@latest succeeds, but dist/gateway.js is missing in the installed package. Steps to reproduce: npm uninstall -g openclaw npm cache clean --force npm install -g openclaw@latest dir "%APPDATA%\npm\node_modules\openclaw\dist\gateway.js" → Path not found openclaw gateway start → Service starts but no node process, port 18789 not listening Browser: http://127.0.0.1:18789/ → ERR_CONNECTION_REFUSED Direct run: node .../dist/gateway.js --port 18789 → Error: Cannot find module '.../dist/gateway.js' Tried: Multiple reinstalls Taobao mirror Deleted ~/.openclaw Task Scheduler task enabled + highest privileges Firewall allows node.exe Disabled antivirus/360 Log snippet: Error: Cannot find module '...\dist\gateway.js' at Module._resolveFilenameName ... Is this a known packaging issue in 2026.3.13? Is there a workaround or fixed version? Thanks for any help!

Qwen3.5 GUFF

[Bug]: microsoft-foundry provider selects Claude deployments but routes them through OpenAI Foundry endpoints

The built-in `microsoft-foundry` provider can discover/select Anthropic Claude deployments from Azure AI Foundry, but it still normalizes them onto the OpenAI-compatible runtime path (`/openai/v1`, `chat/completions` or `responses`) instead of the Anthropic Foundry path (`/anthropic/v1/messages`). This makes Claude deployments appear selectable in onboarding/config, but real requests fail at runtime (commonly 404).

[Bug] Webchat model selector prefixes all models with "anthropic/" regardless of actual provider

The webchat control UI model selector incorrectly prefixes all model IDs with `anthropic/`, ignoring the actual provider name. This causes "model not allowed" errors when switching to non-Anthropic models.

[Bug]: SafeOpenERROR after setting up QMD memory for multi-agent setup

After configuring the QMD memory engine for a multi‑agent setup, any operation that runs the **QMD boot/update pipeline via OpenClaw** fails with: - `qmd boot update failed: SafeOpenError: path is not a regular file under root` - `Memory index failed (<agentId>): path is not a regular file under root` At the same time, QMD works correctly when run **directly** with the documented XDG paths, and `openclaw memory status` reports a valid index under the expected per‑agent state directory. For example, `openclaw memory status` for one agent: ```text Memory Search (assistant) Provider: qmd (requested: qmd) Model: qmd Sources: memory, sessions Indexed: 13/19 files · 13 chunks Dirty: no Store: ~/.openclaw/agents/assistant/qmd/xdg-cache/qmd/index.sqlite Workspace: ~/.openclaw/workspace/assistant By source: memory · 13/13 files · 13 chunks sessions · 0/6 files · 0 chunks Vector: ready Batch: disabled (failures 0/0) ``` This `Store:` path lives under the documented per‑agent state directory (`~/.openclaw/agents/<agentId>/qmd/…`) and I don’t think it should trigger `SafeOpenError`. However, when I run `openclaw memory index --force --verbose`, or when the cron job runs the memory indexer, I consistently get: ```text 🦞 OpenClaw 2026.3.28 (f9b1079) Claws out, commit in—let's ship something mildly responsible. Memory Index (main) Provider: qmd (requested: qmd) Model: qmd Sources: memory (MEMORY.md + ~/.openclaw/workspace/memory/*.md), sessions (~/.openclaw/agents/main/sessions/*.jsonl) 19:39:09+00:00 [memory] qmd boot update failed: SafeOpenError: path is not a regular file under root Memory index failed (main): path is not a regular file under root Memory Index (assistant-1) Provider: qmd (requested: qmd) Model: qmd Sources: memory (MEMORY.md + ~/.openclaw/workspace/assistant-1/memory/*.md), sessions (~/.openclaw/agents/assistant-1/sessions/*.jsonl) 19:39:10+00:00 [memory] qmd boot update failed: SafeOpenError: path is not a regular file under root Memory index failed (assistant-1): path is not a regular file under root Memory Index (assistant-2) Provider: qmd (requested: qmd) Model: qmd Sources: memory (MEMORY.md + ~/.openclaw/workspace/assistant-2/memory/*.md), sessions (~/.openclaw/agents/assistant-2/sessions/*.jsonl) 19:39:10+00:00 [memory] qmd boot update failed: SafeOpenError: path is not a regular file under root Memory index failed (assistant-2): path is not a regular file under root ``` In contrast, running QMD directly inside the same container with the documented env vars works and **does not** raise `SafeOpenError`, for example: ```bash STATE_DIR="${OPENCLAW_STATE_DIR:-$HOME/.openclaw}" export XDG_CONFIG_HOME="$STATE_DIR/agents/assistant/qmd/xdg-config" export XDG_CACHE_HOME="$STATE_DIR/agents/assistant/qmd/xdg-cache" qmd update && qmd embed && qmd status # → succeeds, uses ~/.openclaw/agents/assistant/qmd/xdg-cache/qmd/index.sqlite ``` So the bug appears to be that **OpenClaw’s memory indexer / cron wrapper is invoking QMD with a different (or mismatched) “root” for the safe‑open check than the one that works when QMD is run directly with the documented XDG paths.**

WhatsApp: message tool send fails with 'No active WhatsApp Web listener' while channel is connected and replies work

The `message` tool's `send` action always fails with: ``` Error: No active WhatsApp Web listener (account: default). Start the gateway, then link WhatsApp with: openclaw channels login --channel whatsapp --account default. ``` However: - `openclaw channels status --probe` shows: `enabled, configured, linked, running, connected` - Inbound WhatsApp messages are received and routed to sessions correctly - Normal agent replies are delivered to WhatsApp successfully - Only the `message` tool's proactive send (`action=send, channel=whatsapp, target=<phone>`) fails

[Bug]: `dangerouslyDisableDeviceAuth: true` ignored in 2026.3.11 — Control UI rejects HTTP connections with "device identity required"

After upgrading from 2026.3.8 to 2026.3.11, the Control UI rejects all browser connections over HTTP with `code=1008 reason=device identity required`, even with `gateway.controlUi.dangerouslyDisableDeviceAuth: true`. The flag worked correctly in 2026.3.8. Downgrading to 3.8 restores functionality.

Qwen3.5 0.8b keeps printing 'kv cache removal unsupported, clearing cache and returning inputs for reprocessing" id=0 error="model does not support operation', and finally times out and crashes.

Getting the below error while login to liteLLm admin UI

[Bug]: QMD boot probe triggers llama.cpp build on ARM Linux — searchMode: "search" should skip vector probe

When `memory.backend: "qmd"` is enabled with `searchMode: "search"` (BM25-only), OpenClaw still calls `qmd status` at boot to probe vector capability. On ARM Linux without Vulkan (e.g. Oracle Cloud aarch64), this causes QMD to attempt a full CMake + llama.cpp compile, fail, hit OpenClaw's 5s containment timeout, and leave all agents with `Vector: unavailable` and 0 chunks indexed.