English core of the v0.1 thesis. Design evidence from one private research workstream and a replacement-agent stress test — not a market study, not a protocol spec.
The failure is operational, not mnemonic
When agents replace each other, the usual diagnosis is “the model forgot.” That is the wrong bottleneck. A new agent can be given thousands of lines of history and still not know what actually exists now, what was only planned, what was already written to Git or a database, what was verified, what must not be done again, and why a decision that looks arbitrary from the outside was made.
Control plane, not memory layer
The continuity layer sits between the agent and the real systems of work. It is not a second source of truth in place of Git, a database, a ticket system, or a document store. It keeps pointers, checkpoints, rationale, and recovery state.
- Current state lives in the systems of record.
- Continuity state explains how to resume safely.
- Reasoning history is append-only and loaded lazily.
- Evidence stays where it belongs.
Why the structure is called Talmudic
“Talmudic” here is an epistemic shape, not religious content, layout, or scripture. A decision is not stored as an isolated sentence. The unit keeps: premise → question/gap → test → result → distinction → bounded conclusion → what is still unproven → next step → the source.
Frozen principles
- Systems of record remain authoritative.
- Resume is current state, not history.
- History is append-only.
- Evidence is pointer-based.
- Verification is explicit.
- Unknown remains unknown.
- One next exact action.
- Material writes get In-Flight intent.
- State mismatch forces Recovery.
- Later decisions supersede; they do not erase.
- Load context lazily.
- Documentation overhead must stay cheaper than recovery cost.
The distinguishing property is not “the agent remembers.” The work itself can continue.