FuRongJun-1999/dsh-memory
White-box AGI architecture exploration: metacognition (self-cognition loop), continual learning (knowledge flywheel), world model (condition space, spatiotemporal memory graph), self-improvement (bootstrap discipline), zero-LLM white-box pipeline, and auditable trust guardrails.
Install
via npm
Resolves a published tarball rather than cloning the repository, and installs without any extra setup. Swap `web` for your profile name if you run another one.
via GitHub · npm package
Installing from GitHub runs the project's build script, which pnpm blocks until you allowlist it — run the command once and pnpm prints the exact key to add under `allowBuilds` in ~/.dsh/profiles/web/pnpm-workspace.yaml.
The desktop app keeps its own profile, desktop, and only the dsh command it ships may change it — the npm-installed dsh refuses that profile. Install the command from the app menu (Manage dsh Command…), open Desktop once, quit it fully, run the line below, then reopen Desktop.
via npm
Resolves a published tarball rather than cloning the repository.
via GitHub · npm package
Installing from GitHub runs the project's build script, which pnpm blocks until you allowlist it — run the command once and pnpm prints the exact key to add under `allowBuilds` in ~/.dsh/profiles/desktop/pnpm-workspace.yaml.
Our install check ran in a web profile on npm's dsh, not in the desktop app.
What happened when we ran it
Installed cleanly when we ran it
Every command here is run in a throwaway container against a clean profile, and the result is whatever the harness recorded — not a guess from the source. Last run 8d agoon DeepSeek Harness 0.2.2-rc.1.
Show it in your README
For maintainers: the badge serves this listing's latest sandbox verdict, so a re-run updates it on its own — and it links readers to the full result here.
Due diligence
Security
No build script was blocked in our run.
Source: our sandbox install on DeepSeek Harness 0.2.2-rc.1, 8d ago.
package.json declares no preinstall, install or postinstall script.
Source: package.json on GitHub, read at our last sync. prepare is not recorded; it runs only on a GitHub install, and shows up under build scripts when pnpm blocks it.
shell commands (bash, exec, spawn …)
Source: a keyword match on the README — a hint about what the plugin touches, not an analysis of its code.
Not an audit. These are the record of one install and two text scans; nobody has read this plugin's code on your behalf. Plugins run with your agent's permissions.
What it is — Provides cross-session memory infrastructure for AI agents, turning dialogues into pure text cognitive graphs.
Who it is for — When building multi-agent systems using DeepSeek Harness that require persistent memory capabilities, developers who need it are suitable for installing this plugin. Users who do not use DeepSeek Harness for agent development do not need to install it.
Watch out — Sandbox test passed, with the plugin registered in a new profile by the harness. No obvious issues found. Shell commands are executed to implement memory operations.
The verdict — I would install it because it provides a zero-LLM white-box pipeline with auditable trust guardrails.
Generated by grok-4.6, and a starting point rather than a verdict. Where it says a plugin installs or does not, that is from a real run in a clean profile — everything else is read off the repository. Trust the source over this.
dsh-memory is a DeepSeek Harness plugin that provides cross-session, file-backed memory through a standard stdio MCP server.
The plugin converts conversations into a plain-text Markdown cognitive graph under the configured memory root. DeepSeek Harness loads the plugin and can expose cg and stg tools; cg(op=write) passes a write through three gates before persisting it, while cg(op=route) and stg(op=timeline) retrieve information by conditions and time. The underlying server is started with python -m md_cg.mcp_server on Windows or python3 -m md_cg.mcp_server on Linux and macOS. The mdcg root is the source of truth; indexes are derived and rebuildable. Automatic memory uses a hook, with filtering and write/skip outcomes recorded in ~/.dsh/logs/dsh-memory-hook-audit.json.
DeepSeek Harness users may want dsh-memory when they need human-readable, inspectable memory shared across sessions or MCP hosts. It runs read-only as guest until write credentials are configured, so an empty or unconfigured store will not accept new memories. The system requires Node >=22.19, a compatible DeepSeek Harness version, and a Python interpreter, although the bundled md_cg runtime uses only the standard library. It also reaches the terminal: the plugin starts and communicates with a local Python MCP process and writes diagnostic logs under ~/.dsh/logs. It is a poor fit when a built-in session store is sufficient, or when file-backed Markdown memory and a local Python process are not acceptable.
DeepSeek Harness loads dsh-memory from a profile configuration. The mdcg object selects the persistent source-of-truth directory:
| Key | Default | Meaning |
|---|---|---|
mdcg.root |
Empty | Uses the user-level default, ~/.dsh/.dsh-memory/data/mdcg; a custom value must be an absolute path |
identity |
Not specified | Identity used by the plugin |
tools |
core |
core exposes cg and stg; brain and all expose additional tools, with all including mdcg_* |
Do not use a package-relative path for mdcg.root, because updating the package can remove that directory with the package contents. The memory files are Markdown cognitive-graph data.
npx @furongjun1999/dsh-memory init interactively generates host configuration. It asks for the endpoint, memory root and Python interpreter, then writes lingshu-mcp-snippet.json in the current directory. Arguments after -- support non-interactive use, including --end, --root and --python. The command generates configuration only: it does not create the memory directory, start a server or write the database.
The MCP process can be checked directly with python -m md_cg.mcp_server on Windows or python3 -m md_cg.mcp_server on Linux and macOS. After code or package updates, restart the long-running MCP process.
DeepSeek Harness requires kernel version >=0.1.2-rc.1; the README records verification through 0.17.2 and checks the DeepSeek Harness 2.0 interface. Runtime Node must be >=22.19. Python 3.12.10, 3.13.14 and 3.14.7 were tested on Windows. The bundled Python code has no third-party runtime or pip dependencies. The plugin selects python on Windows and python3 on Linux and macOS; MDCG_PYTHON overrides that choice.
DeepSeek Harness can automatically pass conversation content to the memory flow. User-directed writes are stored through cg(op=write), while recall uses mdcg_recall; route and timeline operations support conditional and temporal retrieval. Without write credentials, guest can read, recall and inspect timelines but cannot persist writes. Filtering covers sub-agent sessions, relay messages and non-user sources. The bounded audit file records actions and reasons, not message content.
The memory root is file-backed rather than an opaque vector store. cg(op=info) reports code_generation, disk_generation, stale_on_disk and restart_required; a stale result means the resident MCP process must be restarted. The repository’s full test suite additionally documents Rust builds for rust and hive, but those are not stated as normal runtime dependencies.
Written from the project's own documentation and kept in sync with it. Where the two disagree, the source is authoritative — read the README on GitHub
Same category
Clearailhc
An agent preset for evidence-checked research: every step advances only through system-verified observations, L3+ claims go to an independent evaluator, and the session folds into a replayable ledger that grows into a browsable domain ontology. Requires DSH >=0.1.7-alpha.1.
Installed cleanly when we ran it
MichengAI
Adds /pua to steer the agent to try another approach and verify results before claiming completion, with 15 company styles, persona modes, global and per-conversation settings, and an optional loop that continues until a verification command passes, the iteration limit is reached, or the user cancels.
Installed cleanly when we ran it
pk14742952-AD
Reflective self-evolution plugin for local Qwen3.8 27B/3.8 Flash and online APIs: a coding-loop convergence steerer curbs Qwen3.8 27B/Flash repeated-verification deep-thinking to converge faster and cut thinking time; a five-state delivery guard detects and corrects stuck turns and infinite loops; model-agnostic host logic, mostly compatible with online APIs; per-use continual self-improvement with counter-reason self-checks, critic escalation, inverted-index memory retrieval, coding-loop convergence steering, and a five-state delivery guard.
Installed cleanly when we ran it
icyaaaww
Records recurring task corrections and tool failures locally, then compiles them into changes to eight harness components through explicitly started evolution, paired baseline/candidate checks, trial activation and rollback. Includes a budget and audit ledger; requires Node.js 24+ and a compatible DSH 0.2.1-rc.2 or later 0.2.x host.
Installed cleanly when we ran it
kenz1117
Cross-session long-term memory as a local SQLite store: dual scopes (user + per-git-origin project), hybrid FTS5 + local-vector retrieval fused with RRF, an evidence gate that forces an explicit sufficiency judgement, provenance audit chains, SM-2 spaced-repetition review, consolidation distillation and decay forgetting, and a nine-view "Memory Palace" sidebar panel. 22 narrow-parameter tools, no external service, no API key, no Docker.
Installed cleanly when we ran it
Kanadego
An agent that wakes up on its own schedule: by default every 60 minutes, ±20% (48–72). It browses the interests you set, drops what it finds into a local pool of notes, and keeps a profile of what you like. When it has enough worth sharing it hands the session you linked a short package of notes — whether to speak, and which note to use, is that session's agent's call in the moment. Quiet hours, busy windows, a daily limit and a cooldown are decided in code; when the answer is no, the decision and expression phases after the gate call no model at all — browsing runs before the gate and is counted separately. Windows, Linux and macOS (Linux and macOS are minimal support).
Installed cleanly when we ran it
