Roarpeng/GraphFlow

A local-first memory-and-context harness: a code knowledge graph plus bounded context compression, installed as a DeepSeek Harness plugin.

GraphFlow bills itself as 'the memory & context harness for coding agents': a local-first code knowledge graph plus bounded context compression, shipped as a DeepSeek Harness plugin. The README frames it against static context injection and plain RAG — each request retrieves only what the current decision needs (graph anchors, compressed summaries, similar past episodes, applicable skills) — and lists its harness primitives as memory (a code graph over 10 tree-sitter grammar languages incl. C/C++, TypeScript/JavaScript via the TypeScript compiler), hooks (outcome auto-capture plus Claude Code and DeepSeek Harness lifecycle glue) and skills (a four-class flywheel with canary validation). The README also documents a memory-poisoning guard: skills imported from external sources are treated as unproven until validated locally.

Memory ★ 17
View on GitHub ↗

Install

dsh plugin --profile web add @roarpeng/graphflow

Copied verbatim from the README's Install section, whose DSH bullet reads 'DSH (DeepSeek Harness) uses the plugin path dsh plugin --profile web add @roarpeng/graphflow — pick one registration path, never both (double loading)'. The README's alternative is dsh plugin --profile web add github:Roarpeng/GraphFlow. A global npm install -g @roarpeng/graphflow is the CLI/MCP route; the README warns to pick exactly one registration path.

Compatibility

The README's development section requires Node.js ≥ 20 and npm ≥ 10. It is pure TypeScript/Node with a CLI, MCP (stdio + HTTP) and a VS Code extension. Local-first: indexing, compression and recall need no API key, and the default semantic-embedding backend downloads its model from huggingface.co once, degrading to a fully offline hash backend. Re-running install is documented as idempotent (no duplicated MCP / skill / rules entries).

Details

FAQ

Does it need an API key?
No. The README says indexing, compression and recall are local-first and need no API key; the default embedding backend downloads its model from huggingface.co once and falls back to a fully offline hash backend.
Can I install both the DSH plugin and the CLI?
No — the README says to pick exactly one registration path, never both, because installing the DSH plugin and the global CLI together double-loads the graph.

Alternatives

diqierjia/StrataGate-AgentMemory · highland0971/dsh-native-memory · Co-Engram/Co-Engram

More plugins in Memory

Browse more in Memory

Guides for Memory plugins