Blog · Comparison
DeepSeek Harness memory: four ways to make your agent remember
August 14, 2026 · dshbase
Out of the box, DeepSeek Harness (dsh) remembers things within a session — its append-only event log and AGENTS.md give it a working short-term memory. The moment you close the tab, that state is gone. That gap is why the memory category is the most crowded corner of the whole plugin directory: a half-dozen plugins all promise the same headline ("long-term memory for dsh") while doing fundamentally different things under the hood. This guide puts the four most representative side by side so you can pick one without installing all of them.
What dsh remembers with no plugin at all
Before you reach for any of these, it's worth being precise about the baseline. dsh's native memory is session-scoped: the append-only event log records every turn, tool call, and permission change, and AGENTS.md holds project-level instructions that get injected into each session. That's genuinely useful — it's why a dsh session is reproducible and auditable — but it is not cross-session memory. Close the tab and the next session starts cold. Every plugin below exists to close precisely that gap, and they differ mainly in how they remember: by citation, by archive, by search, or by shared store.
The four approaches at a glance
- dsh-memory — cited memory built on top of dsh's lossless session log.
- dsh-mnemon — a full Mnemon integration: runtime memory + retrievable archives + supervised memories.
- sage-mem — cross-session memory with Chinese-first full-text search, forked from
claude-mem. - dsh-sgme — an SGME engine bridge for multi-agent shared long-term memory over HTTP.
dsh-memory: citations back to the source event
dsh-memory is the most "dsh-native" of the bunch. Instead of building a separate store, it sits on top of the session log that dsh already writes and distills human-auditable facts, each with a citation back to the exact source event. The selling point is trust: when the agent says "you told me your API key rotates on the 1st," you can click through to the turn where it actually read that, rather than taking a fuzzy vector match on faith. We verified it loads cleanly (test: ok) on dsh 0.1.0-rc.6, and it ships on npm as dsh-memory, so install is a single dsh plugin add dsh-memory.
dsh-mnemon: runtime, archives, and a supervisor
dsh-mnemon is the heavyweight option. It's a deep integration of the Mnemon system, splitting memory into three layers: runtime memory (fast, in-context, for the current task), retrievable archives (long-term storage you can query later), and supervised memories (facts that only get committed after review). If you want a memory system that behaves like an actual knowledge base — with a notion of "draft vs. approved" — rather than a plain notepad, this is the one. It also passed our ok verification on dsh 0.1.0-rc.6, with a few more stars than dsh-memory suggesting it's seen real community use.
sage-mem: Chinese-first search
sage-mem is the pick if your notes, issues, and prompts are in Chinese. It's a cross-session memory plugin forked from claude-mem but re-oriented around Chinese-first full-text search. That's a concrete, not cosmetic, difference: a lot of general-purpose embedding and tokenization layers handle CJK text poorly, so a plugin that optimizes tokenization and recall for Chinese from the start will retrieve your 中文 notes noticeably better. It passed our ok test on dsh 0.1.0-rc.6.
dsh-sgme: memory shared across agents over HTTP
dsh-sgme solves a different problem entirely. Where the other three are for one agent remembering across sessions, dsh-sgme is an SGME memory engine bridge that lets multiple agents share long-term memory over HTTP. If you're running parallel dsh instances — or subagents spawned from a parent session — this is the layer that gives them a common, queryable memory instead of each one carrying its own silo. Star count is still low, but it's the only plugin in the memory category targeting the multi-agent case directly, and it verified ok on dsh 0.1.0-rc.6.
How to choose
The decision tree is short:
- You want auditable, citable memory with minimal setup → dsh-memory.
- You want a real knowledge base with runtime, archive, and review tiers → dsh-mnemon.
- Your content is mostly Chinese and recall quality matters → sage-mem.
- You're running multiple agents / subagents that need shared state → dsh-sgme.
Two more worth knowing before you commit: dsh-plugin-meta-memory (structured long-term memory with brief/full unit pairs and auto-injection) and dsh-plugin-claude-bridge (bring your existing Claude Code memory, skills, and config into dsh with zero migration). Both are in the plugin directory; note that the directory currently lists them with a load-fail status on dsh 0.1.0-rc.6, so treat them as "promising, not yet verified" until that changes.
One final note: there are actually two distinct plugins named "dsh-memory" in the wild — the citation-based one described here, and a separate "Lingshu (灵枢)" plugin published under @furongjun1999/dsh-memory focused on a knowledge flywheel and recursive self-reflection. When you install, check the package name so you get the one you meant. The full, test-verified list — including install commands — is in the plugin directory.