Memory & Contextdsh-plugin

Co-Engram

Co-Engram is a self-evolving team memory system for AI agents and teams. It runs as a native Cordis plugin for DeepSeek Harness, registers 40 tools, injects a memory prompt section, and shares a Markdown-in-Git data repo across Claude Code, OpenClaw, and DSH hosts via a process lock.

Stars
10
Forks
0
Open issues
0
Last push
Aug 20, 2026
Latest release
v0.1.5
24h Growth
+0
7d Growth
+0

Installation

dsh plugin --profile <name> add @co-engram/dsh

Overview

Co-Engram is a self-evolving team memory system for AI agents and teams. It runs as a native Cordis plugin for DeepSeek Harness, registers 40 tools, injects a memory prompt section, and shares a Markdown-in-Git data repo across Claude Code, OpenClaw, and DSH hosts via a process lock.

  • Native Cordis plugin for DeepSeek Harness; install via `dsh plugin add @co-engram/dsh` and a dsh.bundle patch enables zero manual config.
  • Registers 40 tools under bare names and injects a `memory:co-engram` prompt section with live signals such as top tags, skills, and path overview.
  • Self-evolving memory: engrams strengthen with successful use, weaken after failures, and pass through light, deep, and rem maintenance stages.
  • Stores each memory as a Markdown file with YAML frontmatter in Git, using stable ULIDs and independent per-edge synapse files.
  • Deduplicates overlapping captures: `engram_create` returns DUPLICATE or UPDATE, reinforcing or merging the original instead of creating copies.
  • Two-layer proposal filter combines a zero-cost rule prefilter with a necessity evaluator to surface reusable memory candidates.
  • Tiered disclosure via `engram_get` tiers (catalog, digest, content, auto) and `engram_list_paths` directory trees with counts.
  • Shares one data repo across Claude Code, OpenClaw, and DeepSeek Harness through a process lock; host-agnostic `@co-engram/core`.

Related plugins