Overview

QwenPaw separates two memory jobs. Scroll keeps exact conversational evidence—turns, tool calls, results, and sequence data—for reconstructing what happened. ReMe-backed long-term memory extracts selected facts, preferences, decisions, and experience into readable Markdown that can be searched, inspected, corrected, and reused. This chapter explains how automatic search and asynchronous writing fit those roles, why provenance and freshness matter, and why a retrieved note should not override the current instruction or workspace state. Exact history supports reconstruction; long-term memory supports efficient reuse, but neither is infallible.

Learning outcomes

Scroll is the stronger path for reconstructing an exact conversation or tool result; ReMe is a more efficient starting point for reusable knowledge.
Readable Markdown makes long-term memory inspectable and editable, but it does not guarantee truth, provenance, or freshness.
Automatic memory search can add bounded context and background writing can lag behind a reply, so retrieved notes require interpretation.

Keywords

QwenPawScrollReMeMemoryRecallMarkdown

Two kinds of memory: exact history and usable knowledge

QwenPaw has two different memory problems, and solving one does not solve the other. Scroll preserves exact conversational evidence so a long-running task can recover what happened. ReMe-backed memory organizes useful facts, preferences, decisions, and experience into a searchable personal knowledge base. This distinction helps a reader choose the right retrieval path: use durable history when exact reconstruction matters, and use long-term memory when the goal is reusable knowledge rather than a transcript.

A record of events is not yet personal knowledge

The previous chapter treated Scroll's conversation history as an archive. It records turns, tool calls, results, sequence positions, timestamps, and other structure that can support exact recall. Its central question is forensic: what was said, what did the agent call, and what did the tool return? The Scroll index then gives the model a compact directory into that archive.

Long-term memory has a different purpose. A personal assistant should not need to reread every raw conversation whenever a user asks about a standing preference or a decision made months ago. It needs a way to extract durable knowledge from repeated interactions and present that knowledge in a form that can be searched, inspected, corrected, and reused. QwenPaw's ReMe integration treats the workspace as a vault of Markdown memory rather than hiding all extracted knowledge behind an opaque database.

The two layers can refer to the same experience without being interchangeable. Suppose a user once decided that a project should use a particular release process. Exact history can recover the original discussion, the tool calls that inspected the repository, and the evidence available at that time. Long-term memory can preserve the resulting preference or decision as a concise note that helps with a later task. If the user asks, “What did we decide, and why did the test fail that day?” the exact archive is the safer source. If the user asks, “What release process do I usually prefer?” the organized knowledge base is a more efficient starting point.

erDiagram USER_TURN ||--o{ LIVE_CONTEXT : enters LIVE_CONTEXT ||--|| HISTORY_ROW : persists_as HISTORY_ROW ||--o| SCROLL_INDEX : summarized_by SCROLL_INDEX ||--o{ RECALL_QUERY : expands_with USER_TURN ||--o{ MEMORY_NOTE : informs MEMORY_NOTE }o--|| REME_VAULT : stored_in REME_VAULT ||--o{ SEMANTIC_SEARCH : searched_by SEMANTIC_SEARCH }o--o{ MEMORY_NOTE : returns REME_VAULT ||--o{ BACKGROUND_JOB : maintained_by

The entities in this diagram have different authority. A history row is a durable record of an event. A Scroll index entry is a compact navigation aid. A Markdown memory note is an extracted or edited representation of what may be useful later. A semantic search result is a retrieval result, not a new fact. Keeping those roles separate prevents a convenient summary from silently outranking the evidence from which it was derived.

Why make long-term memory readable and editable?

QwenPaw's memory guidance describes a core MEMORY.md, daily notes, and topic-named session notes. The layout is intentionally understandable to a person who opens the workspace.

Published by DeepMox

FAQ

When should I use Scroll instead of ReMe?

Use Scroll when you need exact reconstruction of a turn, tool call, result, or sequence; use ReMe when you need a reusable preference, decision, or other extracted knowledge.

Is a ReMe memory note the original conversation?

No. A ReMe note is a derived or edited representation of useful knowledge, not the original conversation. Inspect the exact Scroll history when the causal chain or wording matters.

When will QwenPaw's long-term memory include a completed chat?

QwenPaw's long-term memory may include a completed chat only after its background extraction and persistence task finishes, so it can lag behind the reply.

Previous9 minQwenPaw Scroll: Managing Long Context

Next9 minQwenPaw Capability Governance

QwenPaw Memory: History and Knowledge

Two kinds of memory: exact history and usable knowledge

A record of events is not yet personal knowledge

Why make long-term memory readable and editable?

FAQ