mindifact
mind + artifact — a versioned, verifiable brain for an agent.
Skills registries distribute modules. A mindifact is the whole
installable mind: the rails an agent follows, the scar tissue it must not
relearn, the verbs it may run inside your application, the memory it carries
forward, and the evals that prove it still works.
Memory layers make an agent better-informed; a mindifact
makes it qualified. The unit is never a fact — it is a
proven capability: everything an agent needs to perform a defined
workflow, installed as one dependency-resolved closure, with the evals
that prove the task can actually be done.
Knowledge that pays for itself — and proves it.
One governed conversation measured at $2.92 where the ungoverned
attempts before it burned ~$17; the staged seal brief targets
$0.30. Not a benchmark claim — a
sealed run you can audit.
install a mind
claude plugin install parkforge-brain@mindifact
mindifact install parkforge-cinematics # or just one domain + its deps
1 · captureruns write dated,
measured memory — free-write, zero ceremony
→
2 · gateThe Factory: an author
drafts the rule, an adversary attacks it; the promotion ledger is the
audit log
→
3 · releaseversioned packs,
per-file provenance, sealed evals — failures published, never
rounded up
The unit is the task, not the fact
How a mindifact differs from a memory layer (mem0-style
fact stores) — complementary layers, different jobs:
| memory layer | mindifact |
| unit of value |
a fact — "user prefers aisle seats" |
a capability — "can build a fenced path,
proven" |
| kind of knowing |
knowing-that (declarative recall) |
knowing-how (workflows, tools, traps, contracts) |
| loading model |
retrieve per query, by similarity — fits
open-ended assistance |
install per task, the whole curated closure —
fits a defined workflow |
| negative knowledge |
no concept of it |
traps: what must NOT be done, with the scar that
taught it |
| composition |
facts do not depend on facts |
capabilities require capabilities — a resolved
dependency graph |
| quality question |
"did I fetch the right fact, fast?" |
"can the agent complete the task — and can you
audit the proof?" |
| in hiring terms |
the employee's notebook |
their training and certification |
Facts still ride along — the memory compartment
carries them in service of the workflow, curated rather than
similarity-ranked. A memory layer could feed a mindifact's capture
stage; the factory decides what becomes law.
Packs
One release train, six packs since v1.6.0: install the
meta for everything, or a member pack for one domain — dependency
resolution order is validated by mindifact graph.
| parkforge-brain | meta-pack — the whole train, self-contained; installs everything below |
| parkforge-characters | Mesh to talking, walking NPC: character-intake and character-dialogue. requires: parkforge-core, unreal-scripting |
| parkforge-cinematics | Cutscene primitives (sequences, spawnable cameras, keying laws) plus the staged cinematic family briefs (stretch-limb, cutscene-grab, lead-npc, cinematic-verification). requires: parkforge-core, unreal-scripting, parkforge-characters |
| parkforge-core | Contracts, traps, governance (brain-maintenance + factory), evals, memory, and the scene-query/world-profile verb base every other pack depends on. |
| parkforge-terrain | Ground, routes, fences: the terrain-strokes skill, the pure-math terrain core, and the terrain/route/fence toolsets. requires: parkforge-core |
| unreal-scripting | Engine-generic Unreal automation: Blueprint graph authoring from Python and fresh-project bootstrap. |
Registry — 19 artifacts, searchable
Everything published across the known sources: mindifact
packs (meshforge namespace) and agent-office skill artifacts
(digest-pinned lockfile). New artifacts appear on the next site build.
| parkforge-brain pack | v1.6.0 | meta-pack — the whole ParkForge train |
| parkforge-characters pack | v1.6.0 | Mesh to talking, walking NPC: character-intake and character-dialogue |
| parkforge-cinematics pack | v1.6.0 | Cutscene primitives (sequences, spawnable cameras, keying laws) plus the staged cinematic family briefs (stretch-limb, cutscene-grab, lead-npc, cinematic-verification) |
| parkforge-core pack | v1.6.0 | Contracts, traps, governance (brain-maintenance + factory), evals, memory, and the scene-query/world-profile verb base every other pack depends on |
| parkforge-terrain pack | v1.6.0 | Ground, routes, fences: the terrain-strokes skill, the pure-math terrain core, and the terrain/route/fence toolsets |
| unreal-scripting pack | v1.6.0 | Engine-generic Unreal automation: Blueprint graph authoring from Python and fresh-project bootstrap |
| cluster-update-advisor skill | v0.1.0 | agent-office skill artifact (OCI, digest-pinned) |
| find-token skill | v0.1.0 | agent-office skill artifact (OCI, digest-pinned) |
| genesis-train skill | v0.1.0 | agent-office skill artifact (OCI, digest-pinned) |
| kubernetes-docs skill | v0.1.0 | agent-office skill artifact (OCI, digest-pinned) |
| mcp-builder skill | v0.1.0 | agent-office skill artifact (OCI, digest-pinned) |
| openshift-docs skill | v0.1.0 | agent-office skill artifact (OCI, digest-pinned) |
| product-lifecycle skill | v0.1.0 | agent-office skill artifact (OCI, digest-pinned) |
| project-board-management skill | v0.1.0 | agent-office skill artifact (OCI, digest-pinned) |
| skill-creator skill | v0.1.0 | agent-office skill artifact (OCI, digest-pinned) |
| weekly-ops-report skill | v0.1.0 | agent-office skill artifact (OCI, digest-pinned) |
| wiki-clip skill | v0.1.0 | agent-office skill artifact (OCI, digest-pinned) |
| wiki-search skill | v0.1.0 | agent-office skill artifact (OCI, digest-pinned) |
| wiki-write skill | v0.1.0 | agent-office skill artifact (OCI, digest-pinned) |
Agent-driven Unreal world-building for welcometotomorrow: terrain strokes, routes, asset/character pipelines, and the operating judgment distilled from the sessions that built them.
| contracts | 2 files | invariants an operator may never violate |
| skills | 6 skills · 23 files | operating rails, open Agent Skills format |
| traps | 17 traps · 1 file | numbered scar tissue, each citing its failure |
| verbs | 32 files | judgment compiled to code, installed into the host app |
| memory | 27 entries | dated observations — the only runtime-writable compartment |
| factory | 24 files | author/adversary briefs that gate every change |
| evals | 9 files | replays that gate the release |
118 files · 7 compartments · every file
sha256'd and git-traced (91/118 from committed sources)
· built 2026-08-10 17:41:48
host application: Unreal Engine 5.8 (welcometotomorrow project, ModelContextProtocol + AllToolsets + ToolsetRegistry plugins enabled)
provides unreal.terrain unreal.routes unreal.viewport unreal.character-intake unreal.character-dialogue
Install the chat half — a real Claude Code plugin
The pack emits a marketplace and a plugin that Anthropic's own tooling
validates and installs. No approval process is involved: a marketplace is
a git repo you host.
claude plugin marketplace add deanpeterson/mindifact
claude plugin install parkforge-brain@mindifact
claude plugin details parkforge-brain
That mounts brain-maintenance, character-dialogue, character-intake, parkforge-boot, terrain-strokes, unreal-blueprint-scripting, unreal-project-bootstrap plus the editor MCP wiring.
Get the CLI
One stdlib Python file, no dependencies, 247 KB.
sha256 ea119b0f4e984b73c8da4bb84d11a00d8a5562763c90a9c45cd7b24bcfbc64c4
curl -fsSL https://mindifact.ai/mindifact -o /usr/local/bin/mindifact
chmod +x /usr/local/bin/mindifact
mindifact setup # pointer-only config; doctor proves every credential it names
mindifact # the guided wizard: pick a pack, answer, confirm
Install the half a plugin cannot carry
A plugin mounts skills into the agent. It has no way to put working code
inside the application the agent operates, and no compartment for memory.
That is the gap a mindifact fills:
python3 brainpack/pack.py install --pack <pack.tar.gz> --unreal <UnrealProject>
python3 brainpack/pack.py install --pack <pack.tar.gz> --workspace ~/AgentWorkspace
The first deploys the verbs into the project's Content/Python,
templates the registration bootstrap, wires the required engine plugins, and
drops a world-profile template. The second writes the boot contract, mounts
the skills, and merges memory without overwriting anything the local agent
has already learned.
Where the line is
This page is generated from the pack itself, so the inventory above cannot
drift from reality. The roadmap below is hand-maintained and deliberately
unflattering — a registry whose pitch is verification has no business
overstating itself.
| proven | Pack build with integrity | 90 files across 7 compartments, every one sha256'd into MANIFEST.json, with per-file git provenance (repo, commit, committed/modified status). |
| proven | Emits a real Claude Code plugin | pack.py plugin writes a marketplace; claude plugin validate passes and claude plugin install parkforge-brain@mindifact installs 7 skills + 1 MCP server, ~1,484 tokens always-on. |
| proven | Installs into a host application | pack.py install --unreal deploys verbs into Content/Python, templates init_unreal.py, wires the .uproject plugins, and drops a world-profile template. This is the part a plugin cannot do. |
| proven | Installs into an agent workspace | pack.py install --workspace writes the boot contract, mounts skills, wires MCP, and MERGES memory without clobbering entries the local agent has already written. |
| proven | Governed promotion loop | Runs record dated learnings; the brain-maintenance skill classifies each one and routes knowledge changes through an author/adversary review gate. promote.py can report and bump versions but cannot write knowledge. |
| proven | Eval 0 — pack health | Read-only check against a live editor: verbs exposed, state stores coherent. Passes from a fresh unpack. |
| proven | Sealed eval runs — RemakeBench-shaped harness | evals/run_eval.py: frozen sha-stamped task prompts (tampering refuses to run), sealed run records (manifest + checksums + transcript), verdicts PASS / FAIL / BLOCKED — never rounded up. First scored run fenced-path-v1 published as-is: mechanical 4 PASS / 2 FAIL (cost $2.32 > $1.50 bar, duration 537s > 420s) / 2 BLOCKED, visual rubric 8/8. The failures are the roadmap: brief-verb-fenced-path-seal targets both. |
| partial | Evals 1, 2 and 4 | Specs written with measurable gates (bridge replay, fresh-world bootstrap, character pipeline) plus a mechanical harness. Not yet run end to end by a fresh agent, so no scores are published here. |
| partial | Portability beyond one project | The world profile removed the hardcoded park constants and the identity proof passed, but only one project has been driven this way so far. |
| proven | Public distribution channels | The marketplace installs from deanpeterson/mindifact (proven with Anthropic's own CLI end to end), and the pack round-trips through Quay as a digest-verified OCI artifact -- byte-identical back. |
| proven | The mindifact CLI | One stdlib file: install/verify/list/search plus a guided wizard that derives its questions from the pack, discovers Unreal projects, remembers answers, and shows the exact command it runs. 199 tests across five suites, two interpreters. |
| partial | Dependency model | requires/provides/conflicts are reserved manifest fields -- parkforge-brain declares five unreal.* authorities -- with a presence check at install. No resolver, no graphs, by design until a second pack exists. |
| partial | Public registry + search | Each pack now gets a generated detail page (versions, bundles, skills, coordinates — see parkforge-brain), still generated from one local pack. Serving many packs, with S3-backed storage and search, is not built. |
| not built | Published eval scores | The verification pitch only becomes real when scores are produced in CI and shown next to the download. Today evals run by hand. |
| not built | Signing and dependencies | No signatures, no pack-to-pack dependency resolution, no conflict detection between packs that teach contradictory rails. |
Source
The pack, the emitter, the evals and this generator live in
https://github.com/deanpeterson/meshforge — currently a private repository, so the link
resolves only for the maintainer. The canonical brain sources sit under
brain/; the marketplace is emitted by
pack.py plugin.