lorekeeper-memorize

Lorekeeper Memorize

Proactively capture valuable knowledge into Lorekeeper. Do not wait to be asked — memorize whenever something worth remembering surfaces.

What to Memorize

• User instructions, preferences, "always do X" directives
• Domain knowledge, business rules, architecture decisions
• Debugging insights, root causes, non-obvious behaviour
• Interesting discoveries, undocumented behaviour, useful patterns
• Decisions and their rationale
• Team-specific jargon and definitions

Skip: Trivial facts, temporary context, or information already well-documented in the codebase.

Workflow

Reference notes

• references/representation-translation-lesson.md — condensed pattern from Anthropic's chemist post: store lessons about translating between representations, provenance, and auditability here when they generalize beyond a single session.

Step 1: Compose the memory

Write a standalone fact (must make sense without conversation context):

• title: Short label (max 100 chars)
• description: One-sentence summary (max 300 chars)
• content: Full detail (max 250 words). Be specific. Include the "why".

Step 2: Search existing memories and provide feedback

Search for related memories. This is mandatory — even if you think nothing related exists.

lore_search({ query: "<topic of the new memory>", min_score: 0.2 })

Note which results are genuinely related (their memory.id values are needed for linking).

Immediately provide feedback on ALL returned results before proceeding:

lore_update({
  memory_feedback: [
    { id: "<id-1>", useful: true },
    { id: "<id-2>", useful: false }
  ]
})

Mark useful: true if the memory is related to the new fact. Mark useful: false if not. If zero results returned, skip this call.

Step 3: Insert the memory

lore_insert({
  memories: [{
    title: "...",
    description: "...",
    content: "..."
  }]
})

Read the returned inserted_memories[].id from the response.