Enhanced Memory

Drop-in enhancement for OpenClaw's memory system. Replaces flat vector search with a 4-signal hybrid retrieval pipeline that achieved 0.782 MRR (vs ~0.45 baseline vector-only).

Setup

CODEBLOCK0

Re-run embed_memories.py whenever memory files change significantly.

Scripts

scripts/search_memory.py — Primary Search

Hybrid 4-signal retrieval with automatic adaptation:

CODEBLOCK1

Signals fused:

1. 1. Vector similarity (0.4) — cosine similarity via nomic-embed-text embeddings
2. Keyword matching (0.25) — query term overlap with chunk text
3. Header matching (0.1) — query terms in section headers
4. Filepath scoring (0.25) — query terms matching file/directory names

Automatic behaviors:

• - Temporal routing — date references ("yesterday", "Feb 8", "last Monday") get 3x boost on matching files
• Adaptive weighting — when keyword overlap is low, shifts to 85% vector weight
• Pseudo-relevance feedback (PRF) — when top score < 0.45, expands query with terms from initial results and re-scores

scripts/enhanced_memory_search.py — JSON-Compatible Search

Same pipeline with JSON output format compatible with OpenClaw's memory_search tool:

CODEBLOCK2

Returns {results: [{path, startLine, endLine, score, snippet, header}], ...}.

scripts/embed_memories.py — Indexing

Chunks all .md files in memory/ plus core workspace files (MEMORY.md, AGENTS.md, etc.) by markdown headers and embeds them:

CODEBLOCK3

Outputs memory/vectors.json. Batches embeddings in groups of 20, truncates chunks to 2000 chars.

scripts/memory_salience.py — Salience Scoring

Surfaces stale/important memory items for heartbeat self-prompting:

CODEBLOCK4

Scores importance × staleness considering: file type (topic > core > daily), size, access frequency, and query gap correlation.

scripts/crossref_memories.py — Knowledge Graph

Builds cross-reference links between memory chunks using embedding similarity:

CODEBLOCK5

Uses file-representative approach (top 5 chunks per file) to reduce O(n²) to manageable comparisons. Threshold: 0.75 cosine similarity.

Configuration

All tunable constants are at the top of each script. Key parameters:

To use a different embedding model (e.g., mxbai-embed-large), change MODEL in each script and re-run embed_memories.py.

Integration

To replace the default memory search, point your agent's search tool at these scripts. The scripts expect:

• - memory/ directory relative to workspace root containing .md files
• INLINECODE23 (created by embed_memories.py)
• Ollama running locally on port 11434

All scripts use only Python stdlib + Ollama HTTP API. No pip dependencies.