co-engram

<h1>Co-Engram: Self-evolving Team Memory</h1> <p><b>Memories that behave like a brain</b> — they strengthen with use, fade when wrong, and verify themselves. Plain Markdown in Git, for AI agents and teams.</p> <p>English | <a href="./README.zh.md">中文</a></p>

Co-Engram is a self-evolving memory system for AI agents and teams. Unlike traditional vector stores that only retrieve, Co-Engram models memory after the brain: engrams strengthen with use, weaken when they fail, consolidate during sleep, and verify themselves through metacognition.

Works with Claude Code (via MCP), OpenClaw (via plugin SDK), and DeepSeek Harness (via native Cordis plugin), with a host-agnostic TypeScript core you can embed anywhere.

Why Co-Engram

Differentiator	What it means
Stable IDs + single-file layout	Every memory is one Markdown file with YAML frontmatter. The engram has a ULID that never changes, so renames, moves, and rewrites don't break references — while content diffs stay clean in Git.
Per-edge synapses	Connections between memories live as independent files keyed by a deterministic hash of (from, to, kind). No duplicate edges, trivial dedupe, and pruning a stale edge is a single file delete.
Self-maintaining	A maintenance engine runs light (RPE-based reinforcement), deep (consolidation + decay), and rem (metacognition upgrade/refute) stages automatically — no manual tagging required.
Two-layer proposal filter	Implicit memory proposals pass through a rule-based prefilter (Layer 1, zero-cost) plus a necessity evaluator (Layer 2 — rule-based by default, optional LLM) — mechanical repetition gets rejected, only genuinely reusable decisions become candidates.
Host-agnostic core	@co-engram/core has zero host dependencies. Same memory, same tools, whether you use Claude Code, OpenClaw, DeepSeek Harness, or your own agent.

Quickstart

Three commands to get Co-Engram working inside Claude Code:

# 1. Install the MCP server globally
npm install -g @co-engram/claude-code

# 2. Initialize the data repo (a separate Git repo, not inside this project)
mkdir -p ~/team-memory && cd ~/team-memory && git init

# 3. Point co-engram at the data repo (writes ~/.co-engram/config.json)
co-engram config data-root $HOME/team-memory

# 4. Wire into Claude Code
claude mcp add co-engram \
  --scope user \
  -- co-engram-mcp

Restart Claude Code, run /mcp in a new session, and you should see the co-engram tools loaded.

…