Vincent van Gogh, Wheat Field with Cypresses (1889), oil on canvas

Governed work and lasting memory for your coding agents.

Your agent is brilliant for exactly one session. AgentHarness governs the work. A second brain keeps what it learns.

scroll ↓

What it does

Speed was never the problem.

An agent writes in minutes; the week you lose comes after: auditing what it guessed, re-explaining what it forgot. Hyperception closes the loop: governed work becomes cited memory, and the next session starts ahead of the last.

the loop
goals agreed with you, testable, before any code builds parallel units on frozen contracts review every finding attacked, not skimmed merges owned to the squash, earned, not assumed ships provisioned and released, then verified live memory decisions kept, each cited to its session recall the next session inherits the last one

Inside the Claude Code your team already runs. No proxy, no hosted service.

install AgentHarness
# inside Claude Code, once /plugin marketplace add SmartStoreAI/agent-harness /plugin install harness@agent-harness

User scope: follows you into every repo you cd into.

install the memory
# via your gh auth gh api .../hyperception/contents/install.sh \ | bash

Lands at ~/.hyperception/vault/, your disk, nowhere else.

The work · AgentHarness

Work that can't skip a step.

One plugin, installed once, carries the whole SDLC into every repo, and its reach runs past the pull request, through provisioning to a verified release. You sit at the top as a veto gate, not an oracle.

harness · orchestrate-sdlc ✓ wave merged
wave› add CSV export
builders3 · own worktrees
review2 rounds · 0 open
result1 PR · squash
Watch it run. The Overseer agrees the North Stars, freezes the contracts, and fans the wave to three Builders in their own worktrees. Every unit clears the gate chain; deep-review converges the branch; the squash is earned, not assumed.
"Read the logs, not the badge."pr-auto-fix
"No hotfix without a reproduced failing test."incident

The memory · a second brain

A brain that sleeps on it.

Your second brain doesn't just store what past sessions decided, it connects those decisions into a structure you can query and traverse. It lives in ~/.hyperception/vault/, a local-first git repo of its own. The model proposes; a deterministic validator disposes; the uncertain waits for your sign-off.

hyp search answers with receipts: the exact session, the exact prompt.

How the memory graph works →

The work compounds, or it resets.
There is no third option.

the loop

The loop · day and night

Step into the mind.

Watch one session move through both organs: a governed wave becomes cited memory, and the next wave starts already knowing it.

AgentHarness · the wave lands

Execute

The Overseer agrees North Stars, freezes contracts, and fans a wave of Builders. Every unit clears the gate chain and lands as a merge-ready PR.

Out comes evidence (commits, decisions, and the reasons behind them), not a chat transcript.

harness
$ harness:pr-to-merged
✓ gates cleared · review converged · CI green
→ PR #46 merged · squash

One organ alone resets every session. Together, they accumulate.

Runs in your stack

Your code and your memory never leave your infrastructure.

No proxy to route your code through, no data to hand over, no black box in the middle. The one thing that does leave: your prompts still reach Anthropic's models through the same Claude Code you already run, with nothing added in between. What Hyperception captures to memory is redacted before it touches disk or any model.

Guarantees

local-first second brain · your disk known secret patterns redacted before disk, and before any model syncs only to your private remote no proxy · no hosted service · no account

How your data stays yours →

The questions a careful CTO asks first.

No. It governs it. Your team keeps its tools, login, and models. The harness changes what happens between your prompt and a verified release, not just the merged PR. Nothing is proxied or hosted, and your code never leaves your infrastructure.
The discipline lives in the pipeline, not in a human remembering to click. You lose the days after the agent "finishes," auditing guesses and re-explaining context. Governed work plus recalled memory is what closes the gap between fast and finished.
AgentHarness runs inside your Claude Code on your machines. Your second brain is a git repo on disk that only syncs to a private remote you point it at. Secrets are redacted before they touch disk and again before any model reads them. Patterns for tokens, keys, cards, and emails run first, and are configurable.
A rules file describes; the harness enforces: hard gates, adversarial review, a state machine that owns the PR to the merge. And most memory APIs are databases behind a vendor: extraction is typically opaque, provenance often thin, updates usually overwriting. This is a typed graph that is also plain markdown on your disk: every fact cites the session it came from, corrections keep the old belief with the reason it changed, and the whole thing syncs with git, not an account.
You could build the files. What you can't self-build is the stream behind them: field-learned fixes shipped as versioned releases, a harness that sharpens on the runs that expose a gap, and a memory design (typed graph, supersede chains, two-pass digest, provenance) that took real iteration. Copying the files gets you a frozen snapshot.
No. It is commercial and licensed per seat. It installs to your machines and runs there, but the product is the update stream, not the inspectable files. Start with a 14-day pilot on one real repo.
We don't publish any. Agent benchmarks are easy to tune and hard to trust. Run the pilot instead: two weeks on a real branch of a real repo, judged by the PRs it sends you.
It escalates on a budget. You are a veto gate, not an oracle: routine calls arrive as evidence-backed rulings with a veto window, direction-changing decisions block on you, and every ruling is logged.

Two weeks on one real repo
beats any benchmark table.

the pilot

The new way of working

Stop reviewing your agent's guesses.

PRs that arrive with their homework shown: goals traced, contracts honored, review survived, merge earned. And a memory that answers "what did we decide, and why" with a citation instead of a guess. Start with a 14-day pilot on one real repo.