Context vs Memory

Snipara separates source truth from durable state. Context is where documents live. Memory is where decisions, learnings, preferences, and short-lived carryover live.

Rule of thumb

Context first. Memory second. Task commit last.

Quick split

You need to know...UseWhy
What does this RFP, spec, PDF, or client doc say?rlm_context_query or rlm_load_documentThose are source documents and current truth.
What are the team standards or reusable business playbooks?rlm_shared_contextShared collections are still document context, not memory.
What did we decide last week about pricing, UX, or architecture?rlm_recallThat is durable memory, not a raw document lookup.
What preferences or learnings should follow the next session?rlm_session_memories / session-bootstrapThat restores durable memory state and optional short carryover.
What should become reusable memory after this task?rlm_end_of_task_commit / task-commitThat persists only durable outcomes instead of whole transcripts.

Context surfaces

  • rlm_context_query for ranked document retrieval across project docs, parsed business files, and linked context.
  • rlm_shared_context for project-linked standards, best practices, and business playbooks.
  • rlm_load_document when you already know the path and need exact source text.
  • rlm_code_* tools for structural code truth that should not be treated as generic document search.

Memory surfaces

  • rlm_recall for past decisions, preferences, validated learnings, and scoped memory questions.
  • rlm_session_memories for session bootstrap.
  • rlm_end_of_task_commit for durable outcomes after work is complete.
  • rlm_remember or rlm_remember_bulk only when you are explicitly storing a durable memory, not uploading source documents.

Default agent workflow

  1. Start with context retrieval.
  2. Pull shared standards only when the task needs them.
  3. Recall memory only for decisions, learnings, preferences, or carryover.
  4. Do the work.
  5. Commit only durable outcomes back into memory.

Anti-patterns

  • Do not use memory as a substitute for reading a spec, PDF, or client file.
  • Do not upload whole source documents into memory.
  • Do not treat shared context collections as memory; they are still documents.
  • Do not persist noisy task transcripts when a compact task summary is enough.

Where this shows up