adoresever/graph-memory
54379Last commit Aug 14, 2026
graph-memory DSH plugin
Graph Memory is a durable, traceable memory plugin for DeepSeek Harness. It stores conversation knowledge as typed nodes (TASK, SKILL, EVENT) and typed edges, enabling cross-session recall through semantic vector search and FTS5 fallback. The plugin reduces token usage by replacing full history replay with a relevant local subgraph.
How to install the graph-memory DSH plugin
dsh plugin --profile web add /absolute/path/to/graph-memory-1.6.0-beta.1.tgzCopying does not run this command. Review the repository and version before installing the graph-memory DSH plugin.
graph-memory DSH plugin data source
graph-memory DSH plugin snapshot date: Aug 18, 2026
discovered
What the graph-memory DSH plugin can do
- Native DSH/Cordis integration via plugin lifecycle, without forking DSH core.
- Automatic cross-session recall: knowledge from previous sessions is injected during Prompt Assembly.
- Explicit memory tools: gm_record, gm_search, gm_status, gm_stats for managing and inspecting memory.
- Optional vector embeddings with support for DashScope, OpenAI, and local providers; falls back to FTS5 without embeddings.
- Community detection, PageRank, and personalized PageRank for ranking relevant memory subgraphs.
Where the graph-memory DSH plugin fits
- Persist knowledge across multiple chat sessions, such as project goals, resolved bugs, and reusable skills.
- Reduce token consumption in long-running conversations by replacing full history with a compact knowledge subgraph.
- Debug and audit memory recall: use gm_status and gm_search to see what was retrieved and why.
- Explicitly record important decisions or fixes with gm_record for later automatic retrieval.
Who the graph-memory DSH plugin is for
- Developers building AI agents on DeepSeek Harness who need long-term memory.
- Users of DSH who want to reduce context window costs while maintaining conversation continuity.
- Researchers testing graph-based memory patterns in agentic workflows.
graph-memory DSH plugin limitations
- Current beta (1.6.0-beta.1) is not published to npm; must be built from source via git clone.
- Requires Node.js 22.19+ or 24+.
- DSH is still in Developer Preview and may introduce breaking changes.
- Vector embeddings require external API keys (e.g., DashScope); without them, recall falls back to FTS5 (no semantic search).
- Pro version with Neo4j and visual graph workbench is not yet available as a DSH Client Plugin.
graph-memory DSH plugin: from the repository README
Quoted from the adoresever/graph-memory README, the upstream source of the graph-memory DSH plugin. Copyright remains with the original authors.
One memory core, native to DeepSeek Harness, with the OpenClaw plugin entry retained. </p> <p align="center"> <a href="README_CN.md">中文</a> · <a href="#core-advantages">Advantages</a> · <a href="#graph-memory-architecture">Architecture</a> · <a href="#install-on-deepseek-harness">DSH Install</a> · <a href="#graph-memory-pro-as-a-dsh-plugin">Pro Plugin</a> · <a href="docs/DSH_NATIVE_PLAN.md">Technical Report (Chinese)</a> </p> Compaction answers “how much of this conversation still fits?” Graph Memory answers “which past knowledge is worth recalling now?” Reusable conversation knowledge becomes typed nodes: - `TASK`: goals, execution, and outcomes; - `SKILL`: validated reusable methods; - `EVENT`: errors, fixes, decisions, changes, and facts. Typed edges such as `USED_SKILL`, `SOLVED_BY`, `REQUIRES`, `PATCHES`, and `CONFLICTS_WITH` preserve relationships. A new question retrieves a relevant local subgraph instead of replaying the complete history. ## Core advantages ### Native host integration - Loaded by the DSH/Cordis plugin lifecycle, not simulated through an MCP side channel. - Integrates Session, Tool, Agent Loop, Prompt Assembly, LLM, and Credentials seams.
Read the full READMERepository license: MIT
graph-memory DSH plugin questions
How do I install Graph Memory into DeepSeek Harness?
Currently the beta is not on npm, so you must build from source. Clone the repo, run 'npm ci', 'npm test', 'npm run build', and 'npm pack' to generate a tarball. Then use 'npx @deepseek-ai/dsh plugin --profile web add /path/to/tarball.tgz' to install it. After installation, verify it's enabled in Settings → Plugins → Plugin list.
Do I need to set up vector embeddings for it to work?
No. Vector embeddings are optional. If you don't configure them, the plugin falls back to FTS5 lexical search, which still provides cross-session recall but without semantic matching. To enable vectors, set environment variables like GRAPH_MEMORY_EMBEDDING_API_KEY, GRAPH_MEMORY_EMBEDDING_BASE_URL, and GRAPH_MEMORY_EMBEDDING_MODEL.
How does Graph Memory reduce token usage?
Instead of dumping the entire conversation history into the prompt, the plugin retrieves only a relevant local subgraph of past knowledge. It uses community detection, PageRank, and traversal to find the most pertinent nodes and edges. This can cut token usage by up to 75% in certain workflows, as shown in the README's benchmark.
Can I manually add or remove knowledge?
Yes. Use the gm_record tool to explicitly persist a TASK, SKILL, or EVENT. Use gm_search to query the graph. Currently there is no gm_delete tool, but you can clear the entire store by deleting the SQLite database file at $DSH_HOME/graph-memory/graph-memory.db.
Is Graph Memory Pro available as a DSH plugin?
Not yet. The Pro version with Neo4j and a visual graph workbench exists in the repository but only as an OpenClaw plugin. It has not been migrated to a DSH Client Plugin. The README mentions that the architecture is reviewed but implementation is not shipped.