tinqiao-oss/engramory
The Engramory curated-memory discipline as an installable plugin ([npm: dsh-engramory](https://www.npmjs.com/package/dsh-engramory)): a deterministic 200-line / 25 KB cap on the `MEMORY.md` index via `ctx.tools.guard()` — growth denied, a shrinking rewrite always passes — plus the protocol registered as a runtime skill. The store is plain markdown, one file per fact, shared with Claude Code, Codex, Kiro, and OpenClaw.
Install
via GitHub · GitHub source
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.
Due diligence
engramory is a DeepSeek Harness plugin that applies a curated, file-based memory protocol to an agent’s local markdown store. It keeps one MEMORY.md index alongside small, human-readable files containing individual facts, with no database, embeddings or server. The protocol is registered as a runtime skill and loaded as standing rules through CLAUDE.md, AGENTS.md or the host’s rules file.
The index is checked through ctx.tools.guard() by a PreToolUse hook for direct Edit, Write and MultiEdit operations. Growth is denied when MEMORY.md would exceed 200 lines or 25 KB; a rewrite that reduces its size is allowed. The discipline also asks the model to deduplicate before writing, update rather than duplicate, remove incorrect facts, and avoid storing information already covered by git, CLAUDE.md or the code. tools/engramory_doctor.py provides a validator.
This suits small-scale, local memory that people can read, edit and diff, particularly where a loaded index is preferable to a database-backed search layer. It is the wrong choice when writes must be globally enforced across shell, MCP, external-editor or sync-client paths: those bypass the hook, and the curation rules remain model-followed. The store assumes one serialised writer. The catalogue flags a terminal surface and credential requirements; the README does not specify which credentials are needed.
Engramory uses plain markdown as its source of truth: one file per fact plus an always-loaded MEMORY.md index. The live store remains git-ignored. The protocol uses four role-based types: user, feedback, project and reference. feedback is procedural memory and requires Why: and How to apply: lines. Its curation contract covers deduplication before writing, updating instead of duplicating, deleting incorrect facts and excluding information already recorded in git, CLAUDE.md or source code.
The index is intended to stay within both limits below; whichever limit is reached first triggers the cap behaviour.
| Limit | Behaviour |
|---|---|
150 lines / 20 KB |
Warning threshold |
200 lines / 25 KB |
Compact or ask before writing |
The hard backstop is a PreToolUse hook using ctx.tools.guard(). It checks direct Edit, Write and MultiEdit operations affecting MEMORY.md. A change that would grow the index beyond either cap is denied. A shrinking or compaction rewrite is allowed.
The repository also provides tools/engramory_doctor.py, described as a validator for checking the memory store. The excerpt does not specify its command-line arguments or output format.
Engramory is designed for local, file-based memory and does not require a database, embeddings or a server. Its rules are loaded through CLAUDE.md, AGENTS.md or the host’s equivalent rules file. The README describes the project as experimental, version 0.8.0, and assumes a single writer with serialised writes.
The hook is not a global write guard. Bash, PowerShell, background Monitor commands, MCP file tools, external editors and sync clients can bypass it. The curation protocol is therefore best-effort and model-followed rather than guaranteed for every task. It is not intended as a mandatory cross-agent memory layer.
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
volcengine
OpenViking memory and context bundle for DeepSeek Harness: pre-step auto-recall and profile injection, session capture, `viking://` URI guarding, and recall/write memory tools backed by an OpenViking server.
No one-line install — this plugin lives inside a larger repository and publishes no npm package.
vectorize-io
Hindsight, agent memory that learns: long-term project memory with auto recall and retain, knowledge pages, deep reflection, and per-repo memory banks.
No one-line install — this plugin lives inside a larger repository and publishes no npm package.
bowenliang123
Context insight panel: see what the model's context window is made of and how it evolves — composition vs. window size, per-request history, compression/injection events, and per-message token stats.
Co-Engram
Self-evolving team memory as plain Markdown in git: a native Cordis plugin that registers 38 bare-name memory tools plus a prompt-signals section re-evaluated at every assembly, with RPE reinforcement, decay and sleep consolidation; shares one data repo with its Claude Code (MCP) and OpenClaw hosts; verified against DSH 0.1.0-rc.6.
omdsh-dev
Cross-agent, local-first persistent memory plugin for DeepSeek Harness (DSH), powered by Mnemon. It shares long-term memory across Mnemon-enabled agents and adds runtime memory, searchable project documents, semantic recall, knowledge graph, and a Sidebar UI.
Aik358
Auto-memory for DSH: three-layer memory (user / project notes / daily logs) with automatic injection, per-turn auto-consolidation, AI greetings, smart search, a calendar view and a settings page, plus inheritance of other AI tools' memories.
