vpromise/dsh-memory

Persistent context across sessions for every agent: a DeepSeek Harness Cordis plugin that captures activity, c

Persistent context across sessions for every agent, inspired by claude-mem but built from scratch on DeepSeek Harness primitives: capture (subscribes to each session's persisted event stream — user prompts, assistant summaries, paired tool calls/results), compress (end of each turn, buffers go to the LLM via ctx.llm.stream asking for JSON [{text, tags}]; failures logged and dropped — compression never blocks the agent loop), store (one JSONL per project, one memory per line with id/text/tags/timestamp/source session; corrupted lines skipped), and inject (in agent/pre-step waterfall, on first step and every N turns, recent + keyword-relevant memories render into a system layer). Model-callable tools memory_search / memory_save / memory_list let agents query and write the same store explicitly.

Coding & Development ★ 2 updated 2026-08-13 ✅ runtime-tested
View on GitHub ↗

Install

dsh plugin --profile web add github:vpromise/dsh-memory

The bundle is committed to the repository, so no build step is needed: dsh plugin --profile web add github:vpromise/dsh-memory (npm dsh-memory 0.1.0 verified on registry — same package), or mount the row manually in $DSH_HOME/profiles/web/cordis.patch.yml (all profiles: $DSH_HOME/cordis.patch.yml). Requirements: a running DeepSeek Harness installation. Zero infrastructure: one JSONL file per project under $DSH_HOME/memory, an in-memory index, and the LLM you already have.

Compatibility

DeepSeek Harness (Cordis plugin); uses native session events, the agent lifecycle waterfall, and the built-in LLM service. No external processes or database servers.

Details

Recent updates

LLM-compressed long-term memory; per-project JSONL store; auto-inject on first step + every N turns; memory_search/save/list tools.

FAQ

How does it store memories?
One JSONL file per project under $DSH_HOME/memory with an in-memory index — no database servers or external processes.
When are memories injected?
In the agent/pre-step waterfall on the first step and every N turns: recent memories plus ones relevant to the current prompt's keywords.
Can agents query it directly?
Yes — memory_search / memory_save / memory_list tools let agents explicitly query and write the same store.

Alternatives

Bleed00/dsh-claude-mem · Scorp1o117/dsh-tdai-memory · freehul/sgme

More plugins in Coding & Development

Browse more in Coding & Development

Guides for Coding & Development plugins