Knowledge Evolution Protocol (KEP)

A portable brain-first contract for AI coding agents. Gives the agent a persistent, searchable brain/ of durable knowledge plus a mandatory retrieve-before step and a capture-after workflow, so it keeps learning about projects, issues, debugging, and patterns across sessions — in every project.

What it does

• Retrieve before any non-trivial task: search the brain for prior knowledge and apply it.
• Capture after learning a durable fact (fix, pattern, gotcha, project knowledge).
• Announce via [brain] markers so a human can verify it's actually being used.
• Guardrails: never store secrets; summarize, don't dump transcripts; keep retrievals fast.

How it works

• Brain = a folder of small Markdown files with YAML frontmatter, organized by namespace (concepts, patterns, references, toolbox, memory).
• Indexing/search = QMD (qmd search BM25, qmd vsearch semantic, qmd query hybrid). QMD also exposes a stdio MCP server (qmd mcp) that agents register as a native search tool.
• Enforcement = an AGENTS.md contract loaded every session (a skill alone can't guarantee retrieval; AGENTS.md is the always-on layer). A v2 plugin can add hard telemetry via tool.execute + brain/_access.log.

Install

Windows PowerShell

Prerequisites: Node.js with npm, and OpenCode. From PowerShell:

git clone https://github.com/programmerlapar/knowledge-evolution-protocol.git
Set-Location knowledge-evolution-protocol
powershell -ExecutionPolicy Bypass -File .\scripts\install.ps1

The PowerShell installer installs qmd with npm when needed, creates the global $HOME\.config\opencode brain, installs the skill and /remember command, and registers the Windows qmd.cmd mcp command in opencode.json. It does not require Git Bash or Python.

If PowerShell blocks script execution, -ExecutionPolicy Bypass applies only to this process. Set KEP_OPENCODE_DIR, KEP_BRAIN_DIR, KEP_QMD_CMD, or KEP_NO_INSTALL=1 in the PowerShell environment before running the script when you need the same overrides as the Bash installer.

Then restart OpenCode.

macOS

Prerequisites: Git, Node.js 22 or newer with npm, and OpenCode. From Terminal, run these commands from the directory where you keep your tools:

git clone https://github.com/programmerlapar/knowledge-evolution-protocol.git
cd knowledge-evolution-protocol
bash scripts/install.sh

The installer installs QMD globally with npm when needed. It creates the brain at ~/.config/opencode/brain, installs the KEP skill and /remember command, and registers the QMD MCP server in your global OpenCode config. It is safe to run again.

If you already have ~/.config/opencode/opencode.jsonc with comments or trailing commas, the installer may warn that it cannot parse the file as strict JSON. Add this entry manually under the top-level mcp object, preserving any existing MCP entries:

"qmd": { "type": "local", "command": ["qmd", "mcp"], "enabled": true }

Restart OpenCode after installation. Verify the setup with:

qmd collection list
grep -n '"qmd"' ~/.config/opencode/opencode.jsonc 2>/dev/null || \
grep -n '"qmd"' ~/.config/opencode/opencode.json

Linux or Git Bash

git clone https://github.com/programmerlapar/knowledge-evolution-protocol.git
cd knowledge-evolution-protocol
bash scripts/install.sh

The installer (idempotent, safe) will:

1. Install qmd if missing (npm i -g qmd).
2. Create ~/.config/opencode/brain/ and register it as the opencode QMD collection.
3. Auto-detect an existing OpenClaw brain (~/.openclaw/workspace/brain) and register it as an openclaw secondary collection (skipped silently if absent).
4. Write the global ~/.config/opencode/AGENTS.md contract (merged, never clobbers).
5. Install the knowledge-evolution-protocol skill and the /remember command.
6. Register the qmd MCP server in opencode.json(c) (with a backup).

Then restart opencode.

Machine-specific: if your qmd needs a wrapper (e.g. a specific Node version or CPU-forced embeddings), set KEP_USE_WRAPPER=1 to use ~/.openclaw/workspace/scripts/qmd-safe.sh, or point KEP_QMD_CMD at your own wrapper. The default is plain qmd for portability.

Usage

• Automatic. The AGENTS.md contract triggers retrieval and capture on non-trivial tasks.
• Manual capture: /remember <topic>: <what you learned>.

Verify it's working by grepping sessions for [brain] markers:

[brain] searched "<topic>" → 3 hits
[brain] applied references/opencode/foo.md
[brain] captured patterns/programming/bar.md (reindexed)

Structure

PROTOCOL.md              # full contract spec
SKILL.md                 # portable agent skill (retrieval + capture + indexing)
AGENTS.md.template       # always-on contract, installed to ~/.config/opencode/AGENTS.md
command/remember.md      # /remember capture command
scripts/install.sh       # one-shot idempotent installer
scripts/install.ps1      # native Windows PowerShell installer
tests/                   # test suite (Tier 1-2 plumbing; Tier 3 e2e gated by RUN_E2E=1)
brain/                   # empty namespace scaffold (your data lives in ~/.config/opencode/brain)

Testing

bash tests/run.sh              # Tier 1 + 2 (plumbing + contract) — no LLM needed
RUN_E2E=1 bash tests/run.sh     # + Tier 3 real opencode session (requires model API, costs tokens)

• Tier 1 verifies install, qmd search/vsearch/query recall, capture→reindex, and the MCP handshake against a scratch brain.
• Tier 2 verifies the installed contract contains the mandatory sections (regression guard).
• Tier 3 runs a real headless opencode session on a fixture Node project, asserting [brain] searched/[brain] applied retrieval and that a resolved issue is captured and reused on recurrence. Tests use KEP_BRAIN_DIR + a scratch brain so your real knowledge is never touched.

Brain root override

Set KEP_BRAIN_DIR to point the brain at another directory (tests, sandboxes, multiple KBs). Keep the opencode QMD collection pointed at it too.

License

MIT