lesliechowsh/dsh-memo

## Core Features
- Memo exposes three tools, memo_search, memo_remember, and memo_stats, for searching past DSH sessions and stored notes from one plugin command.
- memo_search returns matching sessions, snippets, timestamps, match modes, and notes, with optional limits, session filters, time filters, and note tags.
- memo_remember appends durable notes as JSONL records and returns the stored note, while identical text is reported as a duplicate instead of appended again.
- memo_stats provides a corpus overview containing session count, recent sessions, and note count without requiring query parameters.
- The plugin searches the DSH session corpus through the official sessionQuery backend and keeps distilled notes in a local human-readable file.

## Use Cases
- Agents can answer questions about prior conversations by locating sessions where a topic, decision, or phrase previously appeared.
- Agents can store explicit facts, preferences, conventions, or project decisions with memo_remember and retrieve them later through memo_search.
- Agents can narrow historical search by session identifier, epoch-millisecond time boundary, or note tags when the relevant context is partially known.
- Agents can inspect available memory volume and recent activity through memo_stats before issuing more targeted searches.

## Technical Details
- The install command adds the plugin to the DSH web profile and enables the platform session-search index on first use unless a later patch disables it.
- Search processing combines a whole-query phrase step with a weighted token step, deduplicates results, and orders phrase hits before weighted hits.
- Conversation events are indexed into an inverted index persisted at $DSH_HOME/memo/index.json, while notes are stored at $DSH_HOME/memo/notes.jsonl.
- The plugin does not require a vector database, embedding API, external API keys, background indexer, or network calls, and it reports a closed or missing session-query index as an error.

## Notes
- The default in-memory index is rebuilt after a fresh boot, so the first search in a large corpus can be slower than later searches.
- Chinese queries rely on contiguous Han runs as weighted phrases, and word-level search inside a run is unavailable without an upstream tokenizer change.
- Benchmark figures describe session-localization hit rates under lexical retrieval, not end-to-end answer accuracy, and assistant-quoted or preference questions show lower recall.