How OpenClaw Memory Works Now: MMR, Temporal Decay & RAG

OpenClaw's Memory System Got a Major Upgrade

OpenClaw v2026.2.17 shipped significant improvements to memory search: Maximum Marginal Relevance (MMR) re-ranking, temporal decay scoring, and FTS (Full-Text Search) fallback with query expansion. If you use OpenClaw's memory system — storing notes, documents, or conversation history — these changes meaningfully improve the quality of what your agent retrieves when it searches memory.

The Problem With Pure Semantic Search

Standard vector search (embedding similarity) has a well-known limitation: it returns the most similar documents, but not necessarily the most useful ones. If you've saved 20 articles about the same topic, a semantic search query returns 20 very similar results — you get redundancy, not diversity.

Additionally, older memories often become less relevant over time. An article about AI trends from 18 months ago is less useful than one from last week, even if it's semantically very close to your query.

What MMR Does

Maximum Marginal Relevance solves the redundancy problem. Instead of returning the top-N most similar documents, MMR iteratively selects documents that are:

1. Relevant to your query (high semantic similarity)
2. Different from what's already been selected (maximum marginal relevance)

The result: a set of retrieved memories that covers the topic from multiple angles rather than repeating the same information. For research agents and knowledge base queries, this produces noticeably better context for the AI to reason over.

What Temporal Decay Does

Temporal decay adjusts relevance scores based on when a memory was created or last accessed:

memory:
  hybridSearch:
    mmr: true
    temporalDecay: 0.1  # Higher = faster decay; 0 = no decay

With temporal decay enabled, a document from 2 months ago scores slightly lower than an equivalent document from last week. This is the right behavior for most use cases — you generally want your agent to prefer recent information unless the query specifically calls for historical context.

FTS Fallback With Query Expansion

Vector search can fail when queries use terminology that doesn't match your stored documents' embeddings — for example, searching for "ML" when your documents all say "machine learning." The v2026.2.17 update adds:

• FTS fallback — If vector search returns low-confidence results, OpenClaw now falls back to full-text search (BM25) which matches on exact keywords
• Query expansion — The search system expands your query with synonyms and related terms before searching, improving recall for terminology mismatches

Together, these improvements make OpenClaw's memory much more reliable across the diversity of real-world queries.

Configuring the Memory Schema

The v2026.2.17 release also restored the schema help and label metadata for MMR and temporal decay settings, which means openclaw doctor and config validation surfaces now show correct guidance for these options:

memory:
  backend: "lancedb"    # Or "chroma", "qdrant", "qmd"
  hybridSearch:
    mmr: true
    mmrLambda: 0.5      # Balance between relevance (1.0) and diversity (0.0)
    temporalDecay: 0.05  # Gentle decay — recent slightly preferred
    ftsWeight: 0.3       # Weight for FTS vs vector in hybrid scoring

The QMD Backend

If you use the QMD (Quick Memory Database) backend, note that it manages embeddings internally — you don't need to configure a separate embedding provider. The openclaw doctor command was updated to skip embedding-provider warnings when memory.backend is qmd.

Semantic Memory Search Skill

For power users, the Semantic Memory Search skill (recently added to the awesome-openclaw-skills directory) adds vector-powered search to your OpenClaw markdown memory files using Milvus with hybrid retrieval (dense embeddings + BM25). This is useful if you want a separate, high-performance memory search layer alongside the built-in memory system.

Practical Impact: What You'll Notice

After the MMR + temporal decay update, you should notice:

• More diverse context — When your agent searches memory, it retrieves more varied perspectives rather than repetitive similar documents
• Better recency — Recent memories surface more reliably than older ones for time-sensitive queries
• Fewer "I can't find that" responses — FTS fallback catches keyword-based queries that pure vector search would miss
• More coherent long sessions — Agents that rely on memory for continuity produce more consistent responses