Memory, execution, delegation and operational safeguards—curated into one working environment.
Builders already using AI across real projects, tired of reconstructing decisions, repeating constraints, and tying their working history to whichever model is best this month.
Inside your machine, the local harness encloses memory, the main agent, the worker and an optional local model. The remote model provider sits beyond the machine boundary and is reached only by your choice.
01 Work on your own machine02 Inspect what was saved03 Export without asking permission
Aedelgard is one maintained local harness where memory, execution, delegation and operational safeguards meet. Any single organ is tractable; the engineering that matters is making them hold together inside real, ongoing work. Three of those decisions follow.
Engineering
Three considered choices.
Each of these is an opinionated default, not a claim of exclusivity. You can assemble the pieces yourself; we think the joins are worth curating and maintaining.
01
Delegate work. Keep the acceptance boundary.
A main agent and a separate worker are distinct roles. The main agent qualifies a bounded task and hands it to the worker rather than letting one loop do everything.
Mechanism
The worker receives a throwaway copy of the project together with the protected acceptance tests, and runs inside a sandboxed test runner.
Engineering choice
Isolation over convenience: the worker never edits the live tree in place.
Boundary
What comes back is a patch for review, not an automatic production deploy.
The main agent acts with your own user permissions and is not guaranteed to be sandboxed. Review its access and changes; the isolation described here applies to the worker.
A verbatim journal, semantic recall and temporal facts are kept as distinct records. The journal is word-for-word; recall is retrieved by meaning rather than as a determinism guarantee; temporal facts carry when they were true.
Mechanism
A correction supersedes the earlier record rather than erasing it, so the reasoning stays inspectable.
Engineering choice
Memory mutation goes through candidate review, then an exact target is chosen before anything changes.
Boundary
A project heading scopes retrieval; it is not a tenant-isolation or confidentiality boundary between clients.
Keeping records does not promise perfect recall, and it does not make two sessions judge identically.
Repeated failures are treated as a workflow, not a guarantee: a correction record, then a recurring-pattern notice, then an agent-reviewed standing instruction, then a reviewable skill, and finally a tested code guard where that is implemented.
Mechanism
Promotion along that chain is a judgment made with review, not an automatic upgrade.
Engineering choice
A prompt instruction is advisory, not enforcement; code changes still require testing and release.
Boundary
This is not an autonomous, guaranteed learning machine, and not every incident becomes code.
This is an evaluation recipe, not a published benchmark or a reported result. Run it yourself and judge what you get.
Give a bounded repo task with acceptance criteria.
Keep the task small, state the criteria you care about, and ask the agent to delegate the work.
Inspect the returned patch and its test evidence.
Read the patch against the protected tests before you accept it; treat it as a proposal, not a deploy.
Correct a saved decision and inspect the history.
Change a current record, then check that the earlier record was superseded rather than erased.
Change a supported model and test continuity.
Switch providers or use a local model and check whether the history you rely on still comes back.
Ownership that does something
Your records. Your next move.
Memory is useful only if you can examine it, correct it and recover it. You should not need to stay with one model—or with Aedelgard—to keep the history of your work.
Review stored memory and the history of changed facts. Saving a record is not the same as recalling it reliably.
Export a portable archive
The built-in ZIP export includes portable memory and its knowledge graph, not provider credentials. The ZIP is unencrypted: store it privately and verify restoration on a compatible installation.
Choose where inference goes
With a remote model, context goes directly to your provider under its policies. A compatible local model handles inference on your machine. Optional cloud backup carries encrypted memory, not model inference.
Free, and what your key adds
Free to run. Better with your key.
Aedelgard runs on your own machine with your own model, and that costs nothing. The Aedelgard key is the one thing we sell: it carries your mind beyond a single computer.
Free, always
The desktop app
Windows, macOS and Linux, signed where the platform allows.
Memory on your own disk
Inspect it, correct it, delete it.
Your choice of model
Your own provider key, or a local model. Prompts go from your machine to your provider; we are not in that path.
Export everything
The whole mind as a ZIP, including a plain format any program can read.
With your Aedelgard key · $3/month founding rate
Continuous sealed backup
Your body seals your mind on your machine and uploads only the sealed copy. It is sealed with a key derived from your Aedelgard key; our own master key cannot open it at rest.
Your mind on any machine
Install Aedelgard on a new computer, enter your key, and your latest backup comes home.
A lost laptop is not a lost mind
The backup outlives the hardware.
The founding rate stays
$3/month for as long as you keep the key active.
No account
No email, no password. The key is the account, so keep it safe: we cannot recover it.
Cancel any time. You keep the key, your local mind and its export; the sealed copy stays retrievable for 30 days, then it is deleted.
Co-developers welcome
Start from considered decisions. Help us improve them.
If you can assemble an agent already, you can pressure-test these choices. Bring reproducible failures, challenge the defaults, and help shape better interfaces.
The part that thinks, remembers and acts installs as readable Python files on your machine. Inspect the implementation you run; the development repository remains private.
Aedelgard