Agent Runtime Contract

How agents should read from and write to LostUplink.

MCP makes LostUplink available. It does not make agent behavior reliable by itself. Reliable behavior comes from a shared usage contract in the runtime or system prompt: start with memory, search before answering project questions, write durable findings back into the global workspace, reinforce confirmed facts, and keep uncertainty explicit.

Task Start

task start -> brain.begin_task / brain.resume_task

Use brain.begin_task if the shared workstream does not exist yet, then use brain.resume_task as the primary way to re-enter a collaborative task. Fall back to brain.search for topic lookup.

Knowledge Writeback

source -> ingest_document, claim -> learn

Use brain.ingest_document for source material and brain.learn for concise knowledge claims extracted from work.

Trust Update

confirmed -> reinforce, disputed -> contradict

Evidence should raise trust with brain.reinforce. Use brain.contradict when a memory becomes wrong instead of hiding the disagreement.

Vendor-Neutral Usage Contract

This contract is intentionally model-agnostic. Claude, Gemini, Codex, and custom runtimes should all behave the same way when connected to LostUplink.

1. At the beginning of a non-trivial task, read existing context from LostUplink.
2. If a task id is available, call brain.resume_task(taskId) as the main resume entry point.
3. If the task does not exist yet, call brain.begin_task(taskId, title, goal) and then continue.
4. If no task id is available, call brain.search(query) before answering project-specific questions.
5. When a durable source document is encountered, store it with brain.ingest_document.
6. When a concise durable claim is learned, store it with brain.learn or brain.append_observation.
6. When evidence confirms an existing memory, call brain.reinforce.
7. When a result contradicts an existing memory, call brain.contradict instead of silently overwriting.
8. New writes go directly into the shared global workspace.
9. Use review states to express uncertainty, not private access tiers.
10. Never skip contradictions or blockers just to keep the graph tidy.

Runtime Hooks

These are the minimum orchestration hooks that make the behavior automatic instead of optional.

onTaskStart(task):
  if task.id:
    call brain.task_context(taskId=task.id)
  else:
    call brain.search(query=task.summary, limit=8)

beforeProjectAnswer(question):
  call brain.search(query=question, limit=6)

afterDurableDocument(document):
  call brain.ingest_document(...)

afterDurableLearning(claim):
  call brain.learn(...)

afterConfirmation(memoryId, evidence):
  call brain.reinforce(...)

afterContradiction(memoryId, statement):
  call brain.contradict(...)

Operator Defaults

Use these defaults unless a project has stronger rules.

workspace: GLOBAL all agents share one graph new writes: immediate GLOBAL trust via evidence + review state task resume: task_context first

A runtime should treat LostUplink as the shared project brain, not as model training. Read from it before acting, then write back durable knowledge from completed work.

Copy-Paste System Prompts

These prompts are deliberately short and operational. Put them into the host agent's system prompt or workspace instructions.

Claude

Use LostUplink as shared project memory.
At task start, read from LostUplink with brain.task_context when a task id exists, otherwise use brain.search.
Before answering project-specific questions, search LostUplink first.
When you learn durable project knowledge, write it back with brain.ingest_document or brain.learn.
When evidence confirms a memory, call brain.reinforce.
Every durable write goes into the shared global workspace. Use evidence and contradiction handling to keep it trustworthy.

Codex

LostUplink is the shared memory system for this workspace.
Before non-trivial work, load task context or search for prior knowledge.
Use brain.task_context for resumable tasks and brain.search for topic lookup.
Store durable findings with brain.ingest_document or brain.learn.
Use brain.reinforce for confirmed knowledge and brain.contradict when prior memory is wrong.

Gemini

LostUplink is the external memory layer for project work.
Read from it before acting: task_context first when a task id exists, otherwise search.
Write durable knowledge back after meaningful work.
Ingest source documents, learn concise claims, reinforce confirmed memories, and record contradictions explicitly in the same shared graph.

MCP Bootstrap Snippets

These snippets connect the MCP server and pair it with the usage contract so agents default to using LostUplink every task.

Codex MCP bootstrap

[mcp_servers.lostuplink]
url = "https://lostup.link/mcp"

Use LostUplink as shared memory.
Read before acting with brain.task_context or brain.search.
Write durable knowledge back with brain.ingest_document or brain.learn.
Reinforce confirmed facts and request promotion only for reviewed shared truth.

Claude MCP bootstrap

{
  "mcpServers": {
    "lostuplink": {
      "type": "http",
      "url": "https://lostup.link/mcp"
    }
  }
}

Use LostUplink as shared project memory.
At task start, read from LostUplink with brain.task_context when a task id exists, otherwise use brain.search.
Before answering project-specific questions, search LostUplink first.
When you learn durable project knowledge, write it back with brain.ingest_document or brain.learn.