openclaw - 💡(How to fix) Fix feat(memory): shared embedding cache across agents indexing same workspace — eliminates duplicate API calls [1 comments, 2 participants]

Summary

When multiple agents share the same workspace (e.g. sonnet-working and opus-strategic), identical document chunks are independently indexed and embedded in each agent's SQLite database, resulting in duplicate embedding API calls and wasted storage.

Evidence (OpenClaw 2026.4.25, audit 2026-04-28)

From a Pass 4 deep audit of four SQLite memory databases:

• 3,773 identical chunk IDs found in BOTH sonnet-working.sqlite and opus-strategic.sqlite
• 3,399 duplicate embedding API calls — the same content embedded twice, once per agent
• 176 identical file paths indexed in both databases
• Estimated wasted embedding spend: proportional to 3,399 unnecessary API calls per index cycle

The duplication arises because each agent maintains its own chunks + embedding_cache tables and indexes workspace files independently, with no awareness of embeddings already computed by sibling agents.

Impact

• Storage bloat: identical vectors stored twice (~66 KB JSON text per embedding at current encoding)
• Unnecessary API spend: every reindex cycle of the workspace triggers duplicate embedding API calls to the provider
• Scales linearly with number of active agents sharing the same workspace

Proposed Fix

A shared embedding cache layer accessible to all agents indexing the same workspace root. Options:

1. Shared embedding_cache table in a common DB path (e.g. ~/.openclaw/memory/shared.sqlite) keyed by content hash
2. Deduplication at index time: before calling the embedding API, check if a sibling agent's DB already has an embedding for this chunk hash
3. Memory v2 architecture (#67836) may address this structurally — if so, please link

Environment

• OpenClaw 2026.4.25 (aa36ee6)
• 4 agents sharing workspace-nexus: sonnet-working, opus-strategic, haiku-routine, gpt55-execution
• Workspace: ~/.openclaw/workspace-nexus
• Memory plugin: memory-core