Athena-Public: A Local-First Memory and Governance Layer for Any LLM

The problem Athena picks: memory that belongs to a provider instead of to you

The README states the failure mode plainly: you spend months shaping how a hosted assistant responds to you, a model update lands, and the personality resets. Custom instructions stop working. Old conversations become hard to find. Switching to a different vendor means starting from zero. Athena's answer is to move the memory layer onto your own machine as plain Markdown files that you can read, edit, and version-control, and to keep the model as a swappable component. The README's own framing is "Own the state. Rent the intelligence." The target user is someone who already treats their notes and decision history as an asset, works across more than one model, and is comfortable with files and a terminal. It is not aimed at someone who wants the assistant to handle everything invisibly.

What the mechanism actually is: files, a boot budget, and a set of protocols

Three pieces are visible in the material. First, storage: durable context lives in local Markdown rather than in a provider's cloud, which is what makes it inspectable and portable between ChatGPT, Claude, and Gemini. Second, a tiered boot: the README describes roughly 2K tokens for lightweight chat, around 10K for a /start session, and around 20K for /ultrastart, with the claim that 80 to 98 percent of the context window stays free even after 10,000 sessions. That tiering is the real engineering decision here. Loading everything every time would be simpler and would also make the system unusable at scale, so the boot size is a deliberate trade between how much the model knows about you and how much room it has left to think. Third, governance: six constitutional laws, four capability levels, and what the README calls bounded agency. Law #1 is named as the Committee of Seats, and the worked example in the README has Athena refusing a mathematically correct blackjack double because the stake is the user's entire take-home pay, citing a no-irreversible-ruin rule and pointing at examples/protocols/decision/330-economic-expected-value.md. The stated design intent is that personalization alone produces a mirror, and that the disagreement leg is what keeps the mirror leg from becoming an echo chamber. Whether that holds in practice depends entirely on how well the stored context reflects your actual patterns, which is a curation question, not a code question.

Getting it running without guessing at the setup

The repository is Python and MIT-licensed, with the default branch named main. The README links a Quickstart section, a docs/GETTING_STARTED.md file, and an Open in Codespaces badge pointing at codespaces.new/winstonkoh87/Athena-Public, which is the lowest-friction path if you want to look before installing locally. The documented session commands are /start, /ultrastart, and /end; the README describes /end as the loop that keeps memory from decaying and warns that unpruned memory decays like any archive. Beyond those names, the supplied material does not include the install commands, the Python version, the configuration keys, or the directory layout for the memory store, so I am not going to invent them. Read docs/GETTING_STARTED.md and the Safety page before running anything, since a system that reads your personal notes and can act with bounded agency has a wider blast radius than a chat client.

The curation debt is the real cost, not the install

The README is unusually direct about its own weakness: "compounding needs curation. Keep the /end loop running; unpruned memory decays like any archive." That sentence is the honest version of the pitch, and it should be read as a workload estimate rather than a caveat. A memory layer only compounds if what goes in is worth retrieving later, and nothing in the described design decides that for you. The boot tiers help with retrieval cost, but they do not help with retrieval quality. If your notes are inconsistent, the 10K /start boot is 10K tokens of inconsistency. The second limitation is structural: the value proposition is portability across models, so if you never switch providers, you are paying the file-management and protocol-learning cost for a benefit you will not collect. The third is that the governance layer is opinionated by construction. Six laws and four capability levels mean the system will sometimes decline a premise, as the blackjack example shows. That is the stated feature, but it is also a case where the tool is the wrong choice: if you want an assistant that executes whatever framing you hand it, Athena is designed to do the opposite.

How this differs from a plain RAG pipeline or a notes app with a chat box

A conventional retrieval-augmented setup embeds documents, retrieves the nearest chunks, and pastes them into a prompt. It is stateless between calls and has no opinion about what you should do with the retrieved text. Athena keeps the retrieved material in files you own, which is the same substrate, but adds two things a plain RAG pipeline does not have: a boot budget that changes with the task, and a reasoning layer with named protocols that can push back on the question. The repository topics list graphrag, semantic-search, and vector-search alongside memory and local-first, so retrieval machinery is clearly in scope, but the distinguishing claim in the README is not retrieval quality. It is that the assistant reasons differently because of what it knows about you, and that it has the standing to refuse a premise. Compared with a note-taking app that bolts on a chat box, the difference is the other direction: Athena treats the notes as the durable asset and the model as replaceable, which is a stronger commitment than most note apps make.

Version cadence, licence, and what upgrading actually involves

The releases move fast. v9.9.7 was Meta-Awareness Gate v3 on 14 July 2026, v9.9.8 was the Guard-Integrity Sweep on 25 July 2026, and v9.9.9 was a feature port and portfolio refresh dated 10 September 2026. Three releases in roughly two months, with names that describe behaviour changes to the guard and awareness layers rather than pure dependency bumps. For a project whose selling point is a stable memory substrate, that cadence is worth weighing: the files are yours and should survive an upgrade, but the protocols and laws sitting on top of them are moving. Pin a release if you depend on a specific protocol's behaviour. The licence is MIT, which is permissive and places few obligations on how you use or redistribute the code. That says nothing about the content you store in it, and nothing here is legal advice; if your sessions contain material covered by an employer agreement or a data protection regime, the licence on the code does not settle that question.

Who this is for, and the checks to run before you commit

The profile that fits is an engineer or researcher who already maintains notes, switches between models, and wants the switching to be cheap. The profile that does not fit is someone who wants a hosted assistant with no local files, or someone who will not run a session-closing step. The gap between those two is the whole product. Before adopting, verify four things against the repository rather than the pitch: the actual install path and Python requirements in docs/GETTING_STARTED.md, since the supplied material does not state them; the contents of SAFETY.md, because bounded agency still means agency; the protocol files under examples/protocols/decision/, to see whether the reasoning style matches decisions you actually face; and the token figures for your own typical session, because the difference between a 2K chat boot and a 20K /ultrastart boot is the difference between a fast tool and a slow one. The README's own claim, that after 1,900+ sessions the bottleneck shifted to optimising the operator, is the most useful sentence in it. It is a statement about where the effort goes, and it is a warning that the effort does not stop at install.

Editorial conclusion

Adopt Athena if you already keep notes you intend to reuse and you are willing to run a curation loop on your own machine; the /end loop is named in the README as the thing that keeps memory from decaying, and skipping it turns the whole design into an unpruned archive. Do not adopt it if you want a hosted assistant with zero file management, or if you only ever talk to one provider and have no intention of switching. Before committing, read docs/GETTING_STARTED.md and the Safety page, confirm the token budget for your working style by comparing the ~2K chat boot against /start at roughly 10K and /ultrastart at roughly 20K, and check whether the protocol set in examples/protocols covers the decisions you actually make.

Official sources

1. License: MIT
2. Project website
3. README
4. Releases
5. winstonkoh87/Athena-Public on GitHub