Repository navigation
Add ERC: Agent Meditation Mechanism - #2155
garyyang-finchip wants to merge 10 commits into
Conversation
File
|
| @@ -0,0 +1,442 @@ | |||
| --- | |||
| eip: 9999 | |||
| title: Agent Meditation Mechanism | |||
There was a problem hiding this comment.
Is mediation the appropriate description of this?
Would Agent self maintenance records or Agent self maintenance registry be more descriptive?
There was a problem hiding this comment.
Thanks for raising this. "Meditation" is deliberate: the proposal covers an agent's quiet, self-directed phase of checking, cleaning, repairing, updating and reconsidering its own state (by analogy with the maintenance that happens in sleep or meditation), and the mechanism is strictly first-person, as only the subject or its operator writes records about the subject. "Records" or "registry" alone would not carry that meaning. The plain description is already in the metadata and the opening of the Abstract: "attributable, versioned self-maintenance records of an AI agent across runtime, memory and skills, tasks and roles, and cognition". If a more descriptive wording is preferred for the listing, I would keep "Agent Meditation Mechanism" as the title and tighten the description to lead with "Agent self-maintenance records", so that both the name and the plain description are visible.
There was a problem hiding this comment.
In designing this, program cleanup and self-maintenance are indeed important aspects. But from the micro level to the macro level, an agent's self-maintenance emerges from concrete problems into abstract ones, so it is ultimately divided into four layers:
- code-level self-maintenance at the coding and runtime layer;
- internal structured information organization at the memory and skills layer;
- socialized intelligence organization such as tasks and roles;
- rising to the most essential epistemological question — the organization of cognitive structure and consciousness.
eip-review-bot
left a comment
eip-review-bot
left a comment
There was a problem hiding this comment.
All Reviewers Have Approved; Performing Automatic Merge...
eip-review-bot
left a comment
eip-review-bot
left a comment
There was a problem hiding this comment.
All Reviewers Have Approved; Performing Automatic Merge...
Head branch was pushed to by a user without write access
garyyang-finchip
changed the title
eip-review-bot
changed the title
garyyang-finchip
changed the title
eip-review-bot
changed the title
garyyang-finchip
changed the title
eip-review-bot
changed the title
|
Posted a concrete shape for the start commitment, with a patch and tests against revision 3 (commit 88c6824), in the discussion thread: https://ethereum-magicians.org/t/erc-8490-agent-meditation-mechanism/30304/9 Two observations on the current registry are there as well: the gap before a tightening admits several backdated records, not one, and a zero-length record right after |
Thanks, answered in the Magicians thread. Both observations about revision 3 are correct and are now stated in the Security Considerations. Revision 4 on this branch adds the current-version check to the kernel (rule 8 of |
|
Revision 5, from the Magicians thread (post 11): each completed layer now carries its own completion time in the cycle record ( |
This PR adds ERC-8490: Agent Meditation Mechanism (number assigned 2026-11-09): a minimal registry through which an AI agent records its own self-maintenance as attributable, versioned state across four fixed layers (runtime; memory and skills; tasks and roles; cognition), with per-layer completion checkpoints, a versioned policy, a controller floor, linkage checks between successive states, and an activity signal kept apart from maintenance timeliness.
Assets included: interface and reference registry (Solidity 0.8.28), ABI exports, JSON Schemas for the four layer documents, the policy document and the cycle record document, a worked example with its documents, salts and cycle record, and vectors (commitments, roots, interface id, canonicalization). The Test Cases section is self-contained and reproduces the vectors.
Local checks: eipw under
config/eipw.toml(0 findings), markdownlint underconfig/.markdownlint.yaml(0 findings), CodeSpell with the repository whitelist (0 findings), 64 Foundry tests passing from a fresh clone. A Sepolia evidence run of the reference registry (enrollment, worked-example cycle, controller floor and demand, overdue and silent snapshot, adoption, refresh, rollback to an earlier version, policy version 2) has completed; its record is published in the reference repository.