Summary

openclaw memory index fails for all agents with openai embeddings failed: 400 {"error":{"message":"the input length exceeds the context length"}}. The index never builds; zero chunks are stored.

Steps to reproduce

1. Configure memorySearch with provider: "openai", remote.baseUrl: "http://localhost:11434/v1/", remote.apiKey: "ollama", model: "nomic-embed-text"
2. Have one or more memory files larger than ~6,000 tokens (e.g. a daily notes file with multiple appended sessions)
3. Run openclaw memory index or openclaw memory index --force

Expected behavior

Large files should be chunked into ~400 token segments before being sent to the embedding API. The indexer should never send a document larger than the model's context window in a single call.

Actual behavior

All agents fail immediately. openclaw memory status shows Indexed: 0/3563 files · 0 chunks.

OpenClaw version

2026.2.25 / 2026.2.26

Operating system

macOS 15.3 (Darwin 25.3.0, arm64)z

Install method

openclaw update

Logs, screenshots, and evidence

## Environment

- **OS:** macOS 15.3 (Darwin 25.3.0, arm64)
- **Embedding provider:** openai (Ollama OpenAI-compatible endpoint)
- **Embedding model:** nomic-embed-text (via Ollama at localhost:11434)
- **nomic-embed-text context window:** 8,192 tokens
- **Memory sources:** MEMORY.md + memory/*.md + 8 Obsidian extra paths

Impact and severity

High — semantic memory search is completely non-functional

Additional information

Root Cause

OpenClaw appears to send entire files to the embedding API without pre-chunking. When any file exceeds the model's context limit, the entire agent index fails — not just that file. The verbose output shows embeddings: batch start lines, suggesting batching infrastructure exists but chunking is not applied before the API call.

Workaround

Manually split large memory files so no single file exceeds ~5,000 tokens. Not sustainable for daily notes that grow throughout the day.

Suggested Fix

1. Pre-chunk all source files before embedding, regardless of size
2. Expose a chunkSize / maxChunkTokens config option in memorySearch
3. Handle per-file failures gracefully — skip and warn, don't fail the entire agent index