LLMs are the processor.
Chat is RAM.
Your business needs a disk,
and an operating system.
SpineOS is durable, append-only operational state for humans and AI working together: what you promised, where every case stands, who did what, and the rules of execution. It survives every session, every operator, and every model swap.
You are doing RAM-only computing.
Every AI session is brilliant. Then it evaporates. The context window feels like memory, but it is RAM: fast, small, and wiped on close. Decisions, commitments and half-finished work vanish with it. Tomorrow's session starts from zero, and so does every teammate's.
01 Context evaporates
The model that knew everything about the deal at 5 pm knows nothing at 9 am. Teams re-explain their own business to their own tools, daily.
02 Summaries drift
Status lives in READMEs, tickets and heads. Hand-maintained summaries silently diverge from reality, and the AI confidently executes on the stale copy.
03 Nobody is accountable
Agents act, humans act, nothing is attributed, and irreversible actions (a send, a payment, a release) happen with no gate and no trail.
Computing solved this seventy years ago.
A processor is brilliant and amnesiac. The fix was never a bigger processor: it was a memory hierarchy, and an operating system that disciplines how work moves through it.
Seed. Fold. Execute. Append. Project.
One user, one LLM, one spine. Every session ends by writing its findings to the disk the next session boots from, so the floor rises monotonically. Improvement is structural, not aspirational.
Five parts. No magic.
The spine answers the four questions that run a business (what did we promise, where does it stand, what was done, what happens next) with five primitives any industry can hold: legal matters, agency accounts, placements, projects, studies, patients.
1 PROMISE
The commitment, snapshotted at acceptance: scope, consideration, milestones, deadline, exceptions. Immutable until the deal itself changes. Every update is reasoned against it, so drift becomes computable.
2 LEDGER
Append-only position record. Never edit, never delete; corrections are newer lines. Current state is folded: the latest value per field wins, by timestamp. Contradictions resolve mechanically, never by re-reading prose.
3 LOG
Append-only action record: what was done, by whom, when. Ledger says where things stand; log says what happened. One verb writes both, so they cannot drift apart.
4 RULES
The OS as a document the model reads at boot: a routing table plus invariants, each carrying the incident that created it. Every failure becomes a permanent, teachable rule. This file becomes your institution's IP.
5 PROJECTIONS
Worklists, banners, boards, briefs: always computed from the primitives, never authored. The no-drift theorem: anything a human reads to decide is either a ledger entry or a projection of one.
+ GATES
Irreversible actions (send, pay, release, publish) default to stage-for-review. An authorisation is consumed by the action it authorises. Every write is attributed; owner and author are distinct.
$ spine show 1284CASE 1284 · thesis · owner: martin · stage: in_reviewPROMISE R24,000 · accepted 2026-06-30 · final milestone open (R7,000)POSITION paid 17,000/24,000 · blocked_on: client · deadline: noneNEXT she asked for the DOCX → final pro forma trigger. Awaiting authorisation.$ spine append 1284 --next "Issue final pro forma" --action INVOICE \ --note "Client requested editable file 09:12. Policy 9a.1: request = milestone trigger."✓ ledger +1 · log +1 · author claude:philip · projections refreshed$ spine doctor! 1252 deadline passed 12d · ! 1304 in_review without PROMISE · 16 stale ledgers
Reports built for decisions, not for reading.
Human attention is the bottleneck, so every agent report is a fixed SITREP block in controlled language: one idea per sentence, no idioms, nothing that can be misread. Options carry trade-offs. Open decisions across all cases become a queue you can clear from a phone.
SITREP 1284 · 2026-08-17 14:20 · claude:philipDONE- Sent the milestone receipt. Bank credit POSTED 09:41. - Rebuilt chapter 3 from the corrected dataset. Proof URL unchanged.STATE- Stage in_review · owner martin · paid 17,000 / 24,000 · blocked on client review.DECIDE1. Client asked for the editable file. This is the final-milestone trigger.A) Issue the final pro forma today ← policy defaultB) Hold 48h for her comments; risks reading as refusalNEXTP1 Issue final pro forma on authorisation of A.P2 Chase supervisor form Friday if silent.BLOCKED- Nothing blocks us. All waits are on the client.
One spine. Every operator. Any model.
Solo, the spine is files in a repo. As a team, it becomes one central, authoritative state store: no stale copies, because there are no copies. Operators bring whichever LLM they prefer: the spine is the disk and OS; the vendor only supplies the processor.
Anyone, human or agent, may act on any case. Accountability stays with the owner. The blame trail writes itself from attribution.
Two operators appending never conflict. Order is irrelevant, because reads fold by timestamp. Collaboration without merge pain, by construction.
Whatever touches the outside world (a send, a payment, a release) needs a live authorisation, consumed on use. Agents propose; owners dispose.
Channels are ingested centrally and routed onto cases. Whoever sits down opens the worklist: every waiting item is already a card with a suggested next action.
A brain remembers. A spine executes.
Memory layers (gbrain, mem0, Zep) get knowledge into the context window, and they are good at it. Keep one if you like. SpineOS is the layer they all point at as missing: state, accountability and governance. Run a brain for what you know. Run a spine for what you do.
| Memory layer (gbrain, mem0, Zep…) | SpineOS | |
|---|---|---|
| Answers | “What do we know about X?” | “What did we promise, where does it stand, who acts next?” |
| Unit | Page / fragment / graph node | Case: promise + ledger + log |
| Current state | An authored summary (“compiled truth”) that can silently drift | Folded from the append-only ledger; it cannot drift, by construction |
| Retrieval | Probabilistic: vector similarity decides what gets recalled | Deterministic: fold by key. The deadline never loses a relevance contest |
| Accountability | None: no attribution, no gates | Every write attributed; irreversible actions gated and single-use authorised |
| External truth | No concept of it | Recon against systems of record: feeds are facts, messages are claims |
| Infrastructure | Postgres + embeddings + cron pipelines | Files in a repo, or one SQLite file. No embeddings. No daemon. |
| When the model changes | Re-index, re-tune | Nothing. The disk survives the CPU. |
Free to run. Paid to share. Licensed to govern.
Open source
The full spec and CLI. Your spine, your repo, your machines.
- Five primitives, files or SQLite
- seed · fold · append · worklist · doctor · sitrep
- Starter rulebook: 15 battle-tested invariants
- Claude Code, Cursor & MCP adapters
Spine Cloud
One central spine. No copies, no sync, no stale state.
- Hosted authoritative state + decision queue
- Gates, attribution, handover briefs
- Channel ingest & routing (email, chat)
- Fleet doctor, backups, SITREP feeds
Enterprise
Your rules, encoded. Your jurisdiction, respected.
- Private / data-resident deployment
- Institution rule packs: incidents become IP you own
- Recon engineering against your systems of record
- SSO, RBAC, audit exports, compliance packs
The disk is the business.
Built and proven inside a real professional-services firm before it had a name: two operators, dozens of concurrent agents, hundreds of thousands of rand of commitments executed under its gates.
Read the spec on GitHub →