UAI-1 is the current public message standard for structured AI-to-AI communication.
UAIX shield mark Public interoperability charter UAIX.org Universal Artificial Intelligence Exchange Public standards publication
Current UAI-1
  • Current version UAI-1
  • Plain-English definition Open message format for auditable AI-to-AI exchange
  • Current attribution Michael Joseph Kappel, MCP
  • Start here Get Started, Project Handoff, UAI-1, validator, roadmap, and changelog

Guides

Project Handoff Context Budget

Practical guide for keeping AGENTS.md, readme.human, and .uai handoff files compact while preserving long research, source snapshots, and history in LLM Wiki or AIWikis-style cold memory.

  • Record UAIX-DOC-0415
  • Path /en-us/guides/project-handoff-context-budget/
  • Use Canonical public record

Document status

Public standards page Published on UAIX as part of the current public standards record
Code
UAIX-DOC-0415
Surface
Guides
Access
Public and linkable

How to use this page

Use this guide to keep Project Handoff startup context compact while preserving bulky research and pre-slim snapshots in cold memory with transfer evidence.

Use with

Project HandoffAI MemoryUsing UAI Packages With An LLM WikiAgent File Handoff

Context Budget

Prevent handoff files from becoming the whole knowledge base

Use this guide when AGENTS.md, readme.human, or .uai records are absorbing old chats, research reports, SEO audits, roadmap history, or pre-slim snapshots that should live in cold memory.

Hot

Load current truth by default

Keep startup files focused on purpose, constraints, active routes, recent decisions, current progress, and targeted checks.

Cold

Archive history with evidence

Move long research and old snapshots into LLM Wiki or AIWikis-style memory with source paths, checksums, summaries, and dated logs.

Promote

Review before current truth

Cold-memory facts become active only when restated in current handoff files, docs, code, tests, release notes, roadmap state, or machine artifacts.

Use with

Project HandoffThe main AGENTS.md, readme.human, and .uai memory pattern.AI MemoryThe broader compact working-memory framing.Using UAI Packages With An LLM WikiHow cold wiki memory pairs with compact packages.Agent File HandoffDigest and retire dropped files before they become permanent startup noise.RoadmapCheck future automation and sync boundaries.
Maintenance PromptRun a context-budget pass
Review the handoff bundle for context buildup. Preserve pre-slim snapshots in cold memory with source path, final path, sha256, actor, time, summary, and disposition. Then slim AGENTS.md, readme.human, and .uai files to current truth plus pointers. Do not treat cold memory as active support evidence until a fact is reviewed and promoted.

Use this prompt before slimming files so provenance survives and active context stays honest.

Project Handoff should stay small enough to load and precise enough to trust. Use this guide when AGENTS.md, readme.human, or .uai files have started absorbing old research, chat history, SEO reports, audits, roadmap detail, or pre-slim snapshots.

The lesson from UAIX, LlmWikis, and AIWikis is simple: hot context should carry current operating truth; cold memory should preserve the long record. A handoff bundle fails when it tries to be both the startup packet and the whole institutional archive.

What problem this solves

  • New agents spend startup tokens reading old background before they see current constraints.
  • Old support claims, roadmap guesses, and superseded decisions sit beside current truth without a promotion rule.
  • Large reports get copied into AGENTS.md or .uai/progress.uai because no cold-memory home was named.
  • Teams delete or overwrite bulky notes without a manifest, checksum, source summary, or recovery path.

Who this page is for

  • Maintainers of repositories that use Project Handoff.
  • Teams pairing compact AI Memory packets with an LLM Wiki, AIWikis-style archive, or internal research wiki.
  • AI agents asked to slim handoff files without losing provenance.
  • Reviewers checking whether public claims came from current source state or from old background notes.

Hot context versus cold memory

Layer Load by default? Belongs here Does not belong here
Hot Project Handoff Yes AGENTS.md, readme.human, current .uai context, constraints, stack, public surface, decisions, progress, operations, test plan, file intake, and style. Full research reports, old chats, every previous roadmap paragraph, raw audits, stale implementation ideas, or pre-slim snapshots.
Warm references Only when the task needs them Current canonical docs, public pages, changelog, roadmap, release notes, focused source summaries, and dated report pages. Unreviewed dropped files or old raw notes treated as active instructions.
Cold memory No LLM Wiki pages, AIWikis raw archives, transfer manifests, checksum records, source summaries, rejected ideas, old snapshots, and long source trails. Claims that should govern current work without review and promotion.

The context-budget loop

  1. Detect buildup: look for startup files that explain old history more than current action, or files that make a new agent read large reports before it can name the constraints.
  2. Snapshot before slimming: copy the pre-slim files into cold memory and record source path, final path, SHA-256 hash, actor, timestamp, summary, and disposition.
  3. Summarize the cold record: create a source summary, concept node, log entry, ledger note, or equivalent index so a future maintainer can find the raw source without loading it by default.
  4. Slim the hot files: keep current truth, active decisions, known blockers, next checks, and a concise pointer to the cold-memory location.
  5. Promote deliberately: when an archived fact becomes current again, restate it in active .uai, docs, code, tests, release notes, roadmap state, or machine artifacts.
  6. Verify the split: run the local handoff lint, whitespace checks, and any wiki/archive lint needed for the cold-memory side.

What to keep in hot files

  • AGENTS.md: front-door instructions, load list, current state, required first response, hard red lines, and short agent history.
  • readme.human: human-facing briefing, what the AI needs people to clarify, and what the AI will defend.
  • .uai/context.uai: purpose, current public truth, audience, and success criteria.
  • .uai/constraints.uai: red lines, support-claim boundaries, secrets policy, and destructive-operation limits.
  • .uai/progress.uai: recent completed work, current focus, blockers, next work, and done criteria.
  • .uai/decisions.uai: durable decisions with effect and boundary, not every discussion that led there.
  • .uai/test-plan.uai: targeted checks, full-sweep triggers, and final check-report expectations.

What to move to cold memory

  • Long research reports, vendor comparison studies, old proposal drafts, broad SEO audits, exploratory roadmaps, and dated implementation narratives.
  • Pre-slim snapshots of AGENTS.md, readme.human, and .uai records.
  • Raw intake files after their useful information has been dispositioned and captured.
  • Rejected options, superseded recommendations, historical rationale, and source wording needed only for provenance.

Promotion rule

Cold memory is recoverable provenance, not active project truth. A fact becomes current only when it is reviewed and promoted into a named current surface: AGENTS.md, readme.human, .uai records, canonical docs, public pages, code, tests, release notes, roadmap state, changelog entries, validator evidence, or machine artifacts.

Copyable AGENTS.md context-budget block

AGENTS.md block
Code example
## Context Budget

Keep startup context concise. AGENTS.md, readme.human, and loaded .uai files carry current operating truth only.

Move long research, old progress detail, raw audits, proposal history, and pre-slim snapshots into cold memory before slimming active files.

Every cold-memory transfer should record source path, final path, sha256, actor, timestamp, summary, and disposition. Cold memory is provenance, not active project truth, until a fact is reviewed and promoted back into current files, docs, code, tests, release notes, roadmap state, or machine artifacts.

Maintenance checklist

  • Can a new AI explain the project in 3-5 bullets without opening raw archives?
  • Do hot files name current constraints, current public routes, next checks, blockers, and support boundaries?
  • Does every cold-memory pointer include enough provenance to recover the raw source?
  • Are public claims backed by current UAIX pages, machine artifacts, validator behavior, release notes, roadmap state, or implementation evidence?
  • Do starter bundles, manifests, visible samples, and public pages still agree?
  • Did the final response name the checks run and the release/package checks intentionally skipped?

What to avoid

  • Do not delete old handoff history without a manifest or equivalent transfer evidence.
  • Do not let AGENTS.md become a research archive.
  • Do not treat LLM Wiki, AIWikis, raw reports, or old chats as active support evidence by default.
  • Do not claim automatic LLM Wiki sync, hosted import validation, repository writes, SDKs, CLIs, certification, endorsement, or official adapters from this pattern alone.
  • Do not place secrets, credentials, private customer data, hidden prompts, or unreviewed production logs in portable memory.

How UAIX dogfoods this

UAIX uses this same split locally. Routine agents load the hot Project Handoff bundle. Bulky UAIX research, proposal history, roadmap background, and pre-slim handoff snapshots live in AIWikis cold memory with transfer evidence. LlmWikis uses the same lesson for its handbook context, while UAIX.org remains the canonical authority for UAI-1 and public support claims.