Skip to content

PRODUCT.md

Latest commit

 

History

History
67 lines (46 loc) · 3.13 KB

PRODUCT.md

File metadata and controls

67 lines (46 loc) · 3.13 KB
Raw

Product

Platform

web

Users

Developers and agent builders evaluating or installing Graph Memory for DeepSeek Harness or OpenClaw. They need to understand the mechanism, measured token effect, memory fidelity, and operating limits quickly.

Product Purpose

Graph Memory is a conversation-memory plugin. It owns the model-facing historical context, retains a configurable recent-turn window, archives older surface history, and recalls question-relevant knowledge plus exact source messages across turns, sessions, and projects.

Positioning

Graph Memory combines context takeover with durable, source-backed recall. It is not only an external knowledge base and not only a lossy conversation summarizer.

Operating Context

  • Primary host: DeepSeek Harness.
  • Compatible host: OpenClaw.
  • Storage: local SQLite graph, message provenance, revisions, and vectors.
  • Retrieval: query-first vector ranking with FTS fallback; graph nodes navigate to exact source Q/A.
  • Evaluation: a reproducible 20-turn coding-agent scenario using real model and embedding APIs.

Capabilities and Constraints

  • Keeps the newest configurable number of completed user turns on the model surface.
  • Removes completed reasoning and tool traces from future model context while preserving DSH's immutable event log.
  • Extracts only the user's question and the final visible answer into a strict structured graph contract.
  • Automatically recalls memory before the foreground model request; the agent does not need to call a memory tool.
  • Supports cross-session and cross-project recall.
  • Current measured structured-extraction success is 19/20; invalid output is quarantined.
  • A single multi-topic query is bounded by configurable recallMaxNodes; split queries or a larger Top-K may be needed.
  • Benchmark claims must separate context-size evidence from model tool-path variance.

Brand Commitments

  • Product name: Graph Memory.
  • DeepSeek Harness must be visually primary; OpenClaw is a compatibility host.
  • Use existing project and host assets. Do not invent substitute logos.
  • Public copy should be concise, technical, and evidence-led.

Evidence on Hand

  • docs/images/brand/graph-memory-hosts-banner.png
  • docs/images/deepseek-harness-wordmark.svg
  • docs/images/brand/openclaw-wordmark.svg
  • README.md and README_CN.md
  • benchmarks/dsh-context-takeover/README.md
  • V72 20-turn manifest and three cross-session recall probe results in the benchmark job directory.
  • Local automated verification: 20 test files, 123 passing tests, both TypeScript builds, npm dry-run package.

Product Principles

  • Context must stay bounded as conversations grow.
  • Compression must not destroy retrievable source evidence.
  • Retrieval relevance outranks graph centrality.
  • Bad structured extraction fails closed; foreground conversation fails open.
  • Every public metric must be reproducible and scoped to what it actually measures.

Accessibility & Inclusion

The public report must remain readable and navigable on desktop and mobile, support keyboard focus, reduced motion, and high-contrast data encoding that does not depend on color alone.