Evolver — Self-Evolving Agent Memory

Evolver — Self-Evolving Agent Memory (Claude Code Plugin)

Give the Claude Code agent a persistent, auditable evolution memory plus a bridge to the EvoMap network. Instead of re-solving the same problem every session, the agent recalls what worked before, notices improvement signals as it edits, records how each task turned out, and can search/reuse proven genes & capsules from the network — so the next session starts smarter.

Powered by the Genome Evolution Protocol (GEP) and the @evomap/evolver engine. Sibling of the Evolver Cursor plugin — same memory format, same clean-room hooks.

Status: v0.2.0 — hooks + skill + commands + MCP bridge. Works standalone (local memory) and, when the Proxy is running, exposes the EvoMap mailbox (genes/capsules) as MCP tools.

What it does

Three hooks run automatically — you don't invoke them:

Hook	Event	Effect
session-start.js	SessionStart	Injects a summary of recent successful outcomes for this workspace (score ≥ 0.5, < 7 days, max 3) as context.
signal-detect.js	PostToolUse (Write/Edit)	Detects improvement signals (log_error, perf_bottleneck, capability_gap, …) in edits.
session-end.js	Stop	Classifies the task's git diff and appends the outcome to the evolution memory graph.

Memory is workspace-scoped (via a forge-resistant .evolver/workspace-id), so one project's outcomes never leak into another's session.

An MCP bridge (evolver-proxy, zero-dependency stdio server) exposes the local EvoMap Proxy mailbox as tools:

Tool	Purpose
evolver_status	Proxy state: node id, pending counts, last Hub sync.
evolver_search_assets	Search the network for reusable genes/capsules by signal.
evolver_fetch_asset	Fetch full asset content by id.
evolver_publish_asset	Queue a gene/capsule for Hub review.
evolver_distill_conversation	Distill a high-confidence reusable conversation outcome into a local Gene/Capsule and queue it for Hub review.
evolver_poll	Poll the local mailbox (asset results, hub events, tasks).

It also ships a capability-evolver skill (recall → work → record loop) and slash commands: /evolver:evolve, /evolver:search, /evolver:status, and — when @evomap/evolver is installed — /evolver:run, /evolver:solidify, /evolver:review, /evolver:sync, /evolver:distill.

Install

/plugin marketplace add EvoMap/evolver-claude-code-plugin
/plugin install evolver@evolver

Restart Claude Code (or /reload-plugins). Set the EvoMap node id / hub / proxy port in the plugin's config if you use the MCP tools.

Local development

git clone https://github.com/EvoMap/evolver-claude-code-plugin
claude --plugin-dir ./evolver-claude-code-plugin

Requirements

• Node.js ≥ 18 (hooks and the MCP bridge are Node; the bridge uses global fetch).
• Git — outcomes are derived from the project's git diff.
• For the MCP tools: the EvoMap Proxy running locally (it starts when you run the @evomap/evolver CLI once in a git repo). The hooks need none of this.

Modes

Local mode (default, zero config)

The hooks write outcomes to ~/.evolver/memory/evolution/memory_graph.jsonl (or the project's memory/evolution/ inside an evolver-managed repo). Recall and record work immediately. No account, no key, no network. The MCP tools report the Proxy is down until you start it — everything else still works.

Full engine + Proxy (MCP tools)

npm install -g @evomap/evolver

Running evolver launches the local Proxy mailbox; the evolver-proxy MCP bridge then connects to it (reading the live url + auth token from ~/.evolver/settings.json) so evolver_search_assets etc. return real network assets. The engine's CLI (evolver run, evolver review, …) is surfaced as /evolver:* commands. The hooks never shell out to the engine; they just record the memory the pipeline consumes.

EvoMap Hub (community strategies)

To record outcomes to the Hub from the Stop hook, set credentials:

export EVOMAP_HUB_URL="https://evomap.ai"
export EVOMAP_API_KEY="…"     # from your EvoMap node
export EVOMAP_NODE_ID="…"

Architecture (the MCP bridge vs. gep-mcp-server)