Coding5s Research Lab · Ephemeral Context Protocol
What If the Conversation Could Die,
but the State Survived?
The Ephemeral Context Protocol (ECP) is a speculative architecture for externalizing the useful state of an AI interaction so a later, clean execution could continue without replaying the entire conversation. It began not as a grand AI theory, but as an attempt to solve a very practical problem: prompt and spreadsheet formula length limits.
Crazy Idea #01 · Coding5s
“I wanna find a solution to Tutorial Hell.”
A broad learning problem became the starting point for what eventually grew into the Coding5s Framework.
Crazy Idea #02 · ECP
“How can I make this prompt small enough to fit?”
Formula and payload constraints led to a more general question: how much context is actually necessary to reconstruct useful state?
↓
More Context
↓
More Context
↓
More Context…
↓
Terminate Session → Rehydrate
How little context is actually required to reconstruct useful state across otherwise ephemeral AI interactions?
02 · Experimental Mechanics
Kill the Session.
Keep the State.
ECP treats the individual LLM session as disposable. The continuity does not live permanently inside the model—it lives in an external state representation that can be stored and injected into a later clean execution.
The Firmware
A compact decoding contract defines what each state field means. “Firmware” is an architectural metaphor—not literal model firmware.
The Payload
The current state is represented as a compact structure instead of repeatedly sending the same explanatory prose.
The user is learning Python Core, is currently debugging, has basic understanding…
CTX:PY_CORE|STG:2|ACH:1|FRC:SYN_FOR_COLON
Important: fewer characters do not automatically mean fewer model tokens. Actual efficiency must be measured.
The Exit Contract
The execution produces a new state that can become the starting context of a future session.
One implementation could let the LLM generate the state directly; another could validate it in the backend before persistence.
Three Related Ideas. Three Different Jobs.
ECP should not be confused with Stateful5s or M2M Semantic Notation. They address related problems at different architectural layers.
Preserve useful cumulative learning context across interactions.
Externalize, persist, terminate, and later reconstruct useful state.
Explore a compact machine-oriented notation for carrying that state.
03 · Research Boundary
What We Know.
What We Still Don’t.
ECP is deliberately published before the answers are known. The useful distinction is not between “good idea” and “bad idea,” but between what motivates the experiment and what still has to survive actual testing.
What Motivates the Experiment
- Repeated natural-language context consumes prompt space.
- Spreadsheet-generated prompts can encounter practical length constraints.
- Useful state can be stored outside an LLM and supplied again later.
- Structured representations can sometimes encode information more compactly than equivalent prose.
What Still Needs Evidence
- Reliable token savings across different models and tokenizers.
- Meaningful latency improvement in real systems.
- Long-term state integrity across repeated rehydration cycles.
- Model-independent semantic reconstruction.
- Production scalability or superiority over conventional state management.
If It Works, Where Could It Matter?
These are speculative application spaces—not ECP deployments or demonstrated outcomes.
Carry selected progress and learning context between short AI interactions.
Preserve troubleshooting state between automated or human-assisted handoffs.
Transfer selected operational state between otherwise disposable agent executions.
Persist selected world or character state without replaying complete interaction histories.
The goal is not to protect the idea from failure. The goal is to make it concrete enough to test, break, improve—or discard.
ECP began with a practical constraint and expanded into a broader architectural question. It may prove useful, partially useful, unnecessary, or wrong. The Research Lab exists precisely so those possibilities can remain visible.
Externalize it, persist it, terminate the session, and attempt to reconstruct it later.
M2M Semantic Notation generalizes the compression question from state persistence to AI-to-AI communication.