engramory: A Zero-Infrastructure Agent Memory Protocol for DeepSeek Harness

Engramory is a data & knowledge DSH plugin for DeepSeek Harness that turns AI agent long-term memory into a zero-infrastructure, plain-text protocol. It solves the problem of fragmented, hard-to-review agent memory and cross-session inconsistency—without databases, vector embeddings, or servers. Instead, it uses a folder of small, human-readable Markdown files plus one always-loaded index, so memory can be opened, read, edited, and diffed in any editor. Core capabilities include a role/purpose ontology headed by feedback (procedural memory), a curation discipline requiring Why: and How to apply: fields, and an optional hard index cap hook (PreToolUse), enforced by the tools/engramory_doctor.py validator. Maintained by tinqiao-oss under the MIT license, written in Python 3.9+, last updated 2026-08, currently at experimental version 0.8.0.

What core features does the Engramory DSH plugin support?

Engramory is not a new memory architecture but a disciplined superset of the mainstream "markdown files + small index + model curation" pattern. It builds on prior art like Claude Code native auto-memory, basic-memory, obsidian-second-brain, and claude-memory-compiler, but its contribution is the opinionated bundle and discipline: one file equals one fact, the index is always loaded, and feedback type serves as the spine of procedural memory, requiring each memory to state why and how to apply. It also provides a PreToolUse hook that enforces the index cap for direct-edit tools like Edit | Write | MultiEdit, but note it is not a global write guard—shell tools, MCP file tools, external editors, and sync clients can bypass it, so it's best-effort, not guaranteed.

How do I install Engramory?

Installing Engramory as a DSH plugin is straightforward:

dsh plugin --profile web add github:tinqiao-oss/engramory

After installation, the plugin loads as standing rules, e.g., into CLAUDE.md, AGENTS.md, or your host's rules file. AI agents asked to install or check should start at AGENT-SETUP.md, which explains how to determine what your host can enforce, whether a store already exists, what you must not touch, and what to tell the user.

Which DeepSeek Harness versions and platforms are compatible with Engramory?

Engramory is a zero-dependency pure Python tool requiring Python 3.9+, so it is compatible with any DeepSeek Harness environment that can run Python. It is platform-agnostic as long as the host can load standing rules files. Being MIT-licensed, you are free to integrate and modify it. The project is experimental (0.8.0) and assumes a single writer with serialized writes; do not rely on it as a mandatory, reliable, cross-agent memory layer yet.

How do I use Engramory's curation discipline and validator?

The core discipline is "one file = one fact," and each memory must include Why: and How to apply: fields to ensure memories are not just stored but also correctly applied. The validator tools/engramory_doctor.py checks the memory store for compliance, such as index completeness and missing fields. You can integrate it into CI or run it manually to maintain the health of your memory store.

What are the limitations and caveats of Engramory?

The hard index cap hook only applies to matched direct-edit tools and is not a global write guard. Shell tools (Bash, PowerShell), background Monitor commands, MCP file tools, external editors, and sync clients can bypass it. Therefore, the discipline relies on the model following standing rules, making it best-effort. The project assumes a single writer and serialized writes, so it is not suitable for multi-agent concurrent writes. If you need enforced reliable memory across agents, consider waiting for a more mature version.

See more DSH plugins at dsh-plugin.org