Memory
Archive cylinder
Compact semantic notes preserve durable project knowledge while traces remain audit evidence.
Remembering decisions, reloading context, searching prior work, and promoting trace evidence.
The Trace Engine
Fable Harness
A system that teaches AI agents to organize memory context into traces and manageable notes.
Install the Fable-Harness skill (https://github.com/aao-sh/fable-harness) and run it to set up this workspace.Requires Python 3.9 or newer. The npm entry point launches the installer, but Python still runs under the hood.
- Local
- context and memory stay in the workspace
- Traceable
- decisions become reviewable evidence
- Reversible
- rollback plans avoid broad resets
Why use it
Stop asking agents to remember what the repo should prove.
Large agent tasks fall apart when context lives only in chat. Fable Harness turns the workspace itself into the working surface: decisions, notes, traces, checks, and rollback plans live beside the project.
Chat-only memory disappears between sessions.
Unsupported claims creep into progress reports.
Plans drift away from files, tests, and evidence.
Rollback requests become too broad or destructive.
Main workflows
Five calibrated mechanisms for agent work.
Each workflow turns an otherwise fuzzy agent behavior into a visible, reviewable project operation.
Planning
Drafting compass
Plans start from instructions, notes, docs, source files, citations, and graph orientation.
Grounding work before implementation and interviewing the user only when local evidence is not enough.
Decision Loop
Calibrated orrery
Orient, inspect, decide, act, verify, and report with explicit closure checks.
Non-trivial or mutating work where hidden process state would be risky.
Task Parallelism
Split-drive governor
Independent domains can move in waves while dependent loop steps stay ordered.
Complex work across separate files, domains, or responsibilities.
Rollback
Reverse ratchet
Selective rollback plans, checkpoints, backups, and reverse patches keep unrelated work intact.
Undoing a file, hunk, or agent change without erasing unrelated progress.
Evidence loop
A repeatable operating cycle for agent work.
- 01
Orient
Load instructions, memory notes, docs, repo facts, and current state.
- 02
Inspect
Read the files and surfaces that can answer discoverable questions.
- 03
Decide
Choose a scoped next move and record the reason when it matters.
- 04
Act
Change only the relevant surface and preserve unrelated work.
- 05
Verify
Run commands before claiming behavior, coverage, or completion.
- 06
Report
Separate evidence, decisions, limits, and remaining risk.
Community
Join the community and help build open source together.
Product Design
Brief lock, visual alternatives, implementation from a selected direction, and UX/accessibility audit.
Creative Production
Mood boards, logo exploration, original steampunk art direction, and generative polish.
Figma
Componentized mockup, responsive layout review, design tokens, and handoff surfaces.
Canva
Optional derivative social cards or presentation material after the core site is done.
FAQ
Adoption questions.
What does Fable Harness change?
It adds a project-local control layer for agent work: instructions, memory workflows, traces, plans, closure checks, and rollback discipline.
Which agents does it support?
The README describes Codex, Claude Code, and compatible agents, with install targets for Codex, Claude, any AGENTS.md surface, both, or auto detection.
Is Superpowers required?
The installer can skip Superpowers, but the README marks that path as not recommended. The strongest workflow uses both.
How is project-local memory different from global memory?
Global memory can guide general preferences. Project-local memory stores decisions, traces, notes, and retrieval state inside the workspace where future agents can audit it.