Memo

中文文档

Session memory for DeepSeek Harness agents — built directly on the official sessionQuery service, so every session you ever had is searchable memory. Local-first, zero infrastructure, no vector database.

Tools

Tool	What it does
memo_search(query)	Search every past session in this workspace plus your memo notes — snippets, titles, time filtering
memo_remember(text, tags)	Write one durable note: facts, decisions, preferences that survive across sessions
memo_stats()	Corpus overview: session count, recent titles, note count

Install

dsh plugin --profile web add dsh-memo@latest

Restart dsh web. The three memo_* tools appear in your agent's tool list.

Manual profile install (deployments without the dsh plugin subcommand):

1. cd "$DSH_HOME/profiles/web" && npm install dsh-memo

2. Append to the profile's cordis.patch.yml:

   - insert:
       - id: memo
         name: 'dsh-memo'
   

3. Restart dsh web.

Where notes live

memo_remember appends to $DSH_HOME/memo/notes.jsonl — plain JSONL you can edit, back up, or delete freely.

Design & research grounding

Memo sits cleanly on the memory taxonomy of Memory for Large Language Models (Zhoubian, Zhang, Kharlamov & Tang — THUNLP · Tsinghua / NUS), which characterizes memory along three orthogonal axes:

Axis	Memo
Representation	Explicit — independently addressable JSONL logs and notes, decoupled from model computation
Update dynamics	Online — DSH appends every message, tool call, and result as it happens; memo_remember writes distilled notes
Persistence	Long-term — survives context windows, sessions, and process restarts

Writing (memo_remember) and reading (memo_search retrieval with snippets, titles, and time filters) follow the survey's memory-operation view; consolidation and compression are the next milestone.

Benchmark targets

LongMemEval-S (retrieval hit@k/MRR) primary, LoCoMo secondary. Results will be published in this README once available.

License

MIT — see LICENSE.