Link Memories

Use lore_recommend_links to find high-confidence link candidates between memories, then confirm them via lore_insert.

When to Run

• After inserting a batch of memories — new memories have few or zero links
• After a session — lore_reflect + lore_insert creates new facts; follow up with lore_recommend_links on key new memories
• Periodic maintenance — run lore_recommend_links on orphaned memories (those with links: [] in search results)

How to Use

candidates = mcp_lore_recommend_links(lore_id="<memory-id>")

Optional top_k parameter overrides the max candidates returned.

Reading the Output

Each candidate has:

• weighted_score (0.0–1.0): combined score from all 4 signals (cosine, BM25, entity overlap, temporal proximity)
• scores: per-signal breakdown for transparency

The agent evaluates candidates itself — it already has an LLM. lore_recommend_links only surfaces the data.

What Makes a Good Link

• Shared topic/entities: memories about the same person, project, concept, or codebase
• Causal or structural: depends_on for dependencies, references for generic relations
• Chronological: supersedes for newer versions replacing older ones
• Conflict: contradicts when two memories make conflicting claims about the same thing

What to Skip

• Low weighted_score (< 0.3): weak signals, likely spurious
• Temporal-only matches: memories created close in time but about completely different topics

Confirming Links

lore_recommend_links never writes anything. To create a link:

mcp_lore_insert(
    links=[{
        "source_memory_id": "<source-id>",
        "target_memory_id": "<target-id>",
        "relation_type": "references",
        "reason": "similar topic about Python web frameworks"
    }]
)

Relation Types