StrataGate plugin
A DeepSeek Harness plugin: recent talks stay detailed, older ones compress, and key facts live as events with sources.
- 115 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The plugin stores your conversations and tool records, and some text may go to a configured embedding model.
Install
Manual install
dsh plugin --profile web add stratagate-dshThe command from the README quick start: replace web with your DeepSeek Harness profile name.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The plugin stores your conversations and tool records, and some text may go to a configured embedding model.
This is third-party code. Review the repository files before installing.
What it does
StrataGate is a cross-session memory plugin for DeepSeek Harness. It keeps daily notes for quickly resuming recent work and stores each conversation block in six views, L0 to L5, from an index to original messages. Older blocks display in compact form and expand when needed, while full records are preserved. Long-term memory holds important decisions, preferences and requirements as events with scope, time and source references. Related events fold into a topic directory, and an optional knowledge graph derives current information about people, projects and tools from them. Before answering, the agent assesses whether retrieved evidence is sufficient and searches further if not. Events that were actually used raise memory weight, while mere search hits do not. Memory from other AIs can be imported as events with the original text kept.
Who it is for. DeepSeek Harness users who want memory with traceable sources and decay instead of re-explaining the project in every session.
Good fit when
- A new session should pick up progress, blockers and next steps from the three most recent days.
- A month later you must understand why the project moved from npm to pnpm and what was decided earlier.
- Memories from another AI must be imported with the original wording kept.
- The topic directory helps find something when you recall a decision only vaguely.
Not a fit when
- You work in Claude Code, Cursor or Codex: the plugin is written for DeepSeek Harness.
- Conversation history must not be stored even locally.
- A project rules file is enough: layered memory is overkill.
Example request
Recall why we chose pnpm and show the original messages this comes from.Limitations
The plugin targets DeepSeek Harness and is installed into a chosen profile, with supported versions and diagnostics described in docs/DSH.md. Automatic memory recall is capped at four events and about 900 tokens, and the topic directory gets a separate budget of about 400 tokens. Semantic search needs an embedding model configured, and reranking needs a separate model. The knowledge graph is off by default. The author reports evaluation results on a single LoCoMo conversation, so the sample is narrow.
How to disable. Remove the plugin with dsh plugin remove stratagate-dsh for your profile and restart it. Local memory files are removed separately.
FAQ
How do I install StrataGate?
Run `dsh plugin --profile web add stratagate-dsh`, replacing `web` with your profile name, and restart the profile.
Where do I view the memory?
In DSH settings under StrataGate-AgentMemory.
What are the L0 to L5 levels?
Six views of one conversation block, from a title and key facts to full original messages and tool records.
Is an embedding model required?
No, keyword search works without it. An embedding model adds semantically similar events.
Can I bring memory from another AI?
Yes, imported text becomes events with the original kept, as the external import section describes.
Discussion
No questions or reviews yet.
Related
PluginsA DeepSeek Harness plugin that keeps notes between sessions in local SQLite, each write approved and logged in an audit trail
Long-term memory for DeepSeek Harness with an MCP server: the same memories reachable from Claude Code, Cursor and Codex
A shared local SQLite memory for Claude Code, Codex and Hermes: every fact has a source and can be corrected or deleted
Shared semantic memory for agents: conversations are saved as markdown and recalled by search across Claude Code, Codex and other agents
A plugin with memory of past chats: it searches by meaning for decisions and reasons across Claude Code, Codex, Cursor and opencode sessions over MCP
A DeepSeek Harness plugin for a standing team of agents with memory, projects, task owners and verification by a second member
README in short
The documentation explains memory by layers: daily notes, conversation blocks, events, a topic directory and an optional graph. The repository includes LoCoMo evaluation materials, architecture notes and a DSH guide in two languages. A CodeBuddy plugin manifest is also present.