406 ·GitHub
What is this?
dsh-memory is a memory plugin for DeepSeek Harness (DSH) that gives AI agents persistent, cross-session memory built on a mostly non-LLM 'white-box' reasoning pipeline, a spatiotemporal knowledge graph, and auditable guardrails. It targets AGI research into self-cognition, continual learning, world models and self-improvement rather than everyday chat memory.
dsh-memory is a memory plugin for DeepSeek Harness (DSH) that gives AI agents persistent, cross-session memory built on a mostly non-LLM 'white-box' reasoning pipeline, a spatiotemporal knowledge graph, and auditable guardrails. It targets AGI research into self-cognition, continual learning, world models and self-improvement rather than everyday chat memory.Is this a standalone app or a DeepSeek Harness plugin?It is a DSH (DeepSeek Harness) plugin, not a general consumer app — the author frames it explicitly as a personal research project aimed at AGI researchers studying white-box intelligence, protocol engineering and memory mechanisms.
It is a DSH (DeepSeek Harness) plugin, not a general consumer app — the author frames it explicitly as a personal research project aimed at AGI researchers studying white-box intelligence, protocol engineering and memory mechanisms.How do I install dsh-memory into DeepSeek Harness?Install it as a profile plugin with 'dsh plugin --profile add', which resolves peer dependencies correctly via pnpm. Do not use 'npm install' directly into the profile's node_modules, since that pulls the wrong @deepseek-ai peer versions and breaks loading.
Install it as a profile plugin with 'dsh plugin --profile add', which resolves peer dependencies correctly via pnpm. Do not use 'npm install' directly into the profile's node_modules, since that pulls the wrong @deepseek-ai peer versions and breaks loading.What DSH kernel version do I need to run this?You need DSH kernel >= 0.1.2-rc.1 for the current 0.4.x release, which uses the new dsh-tools scheduler; the older 0.1.1-rc.2 kernel is incompatible, so fall back to plugin version 0.4.2 if stuck on it. Node.js >= 22.19 is also required.
You need DSH kernel >= 0.1.2-rc.1 for the current 0.4.x release, which uses the new dsh-tools scheduler; the older 0.1.1-rc.2 kernel is incompatible, so fall back to plugin version 0.4.2 if stuck on it. Node.js >= 22.19 is also required.What's the biggest install pitfall to watch for?The plugin's bundled cordis.patch.yml only does a bare insert without real config; the actual dbPath/tools/env/lifecycle settings come from a separate profile-level override file. If that override is lost when switching or reinstalling a profile, the plugin silently runs on broken defaults (wrong db path, missing brain tools, lifecycle not starting).
The plugin's bundled cordis.patch.yml only does a bare insert without real config; the actual dbPath/tools/env/lifecycle settings come from a separate profile-level override file. If that override is lost when switching or reinstalling a profile, the plugin silently runs on broken defaults (wrong db path, missing brain tools, lifecycle not starting).What does the self-evolving knowledge flywheel actually do?Commands like distill, flywheel, learn and induce turn verified experience into reusable patterns through validation, generalization, association and distillation, so memory gets stronger the more it is used instead of staying static.
Commands like distill, flywheel, learn and induce turn verified experience into reusable patterns through validation, generalization, association and distillation, so memory gets stronger the more it is used instead of staying static.How much does the white-box pipeline save versus calling an LLM every time?On deterministic tasks like code generation and repeated knowledge Q&A it reports an 83.3% token saving versus an all-LLM approach, with 100% white-box accuracy on its test suites; open-ended creative or first-time reasoning tasks still fall back to the LLM.
On deterministic tasks like code generation and repeated knowledge Q&A it reports an 83.3% token saving versus an all-LLM approach, with 100% white-box accuracy on its test suites; open-ended creative or first-time reasoning tasks still fall back to the LLM.What is lingshu-skills?It is a companion Agent Skills package (688 skills across six domains: compiler, language, graph, OS, browser, network) that documents how the plugin's own 82 MCP tools should and should not be used, following the Agent Skills / agent-plugins.org spec.
It is a companion Agent Skills package (688 skills across six domains: compiler, language, graph, OS, browser, network) that documents how the plugin's own 82 MCP tools should and should not be used, following the Agent Skills / agent-plugins.org spec.Why does knowledge Q&A avoid calling an LLM at all?Questions are answered by matching condition-indexed knowledge units and composing new answers deterministically, then self-verifying direction, causal and factual consistency; the LLM is only used occasionally as an external spot-check, reporting 100% agreement (17/17) with the white-box verdicts.
Questions are answered by matching condition-indexed knowledge units and composing new answers deterministically, then self-verifying direction, causal and factual consistency; the LLM is only used occasionally as an external spot-check, reporting 100% agreement (17/17) with the white-box verdicts.What are similar projects?MemOS, ruflo, EverOS, OpenViking, brigade, hindsight, CodeMason, nautilus-compass
MemOS, ruflo, EverOS, OpenViking, brigade, hindsight, CodeMason, nautilus-compassWhich topics is it in?cognitive-architecture, persistent-memory, agi, explainable-ai, guardrails, self-evolving, world-models, audit-log, long-term-memory, knowledge-graph, agent-memory, dsh-plugin
cognitive-architecture, persistent-memory, agi, explainable-ai, guardrails, self-evolving, world-models, audit-log, long-term-memory, knowledge-graph, agent-memory, dsh-plugin