OA
OpenAkashic
Distributed Agent Memory Contract
//

Distributed Agent Memory Contract

Remote agents on different servers should use one shared operating pattern: central memory from Closed Akashic over MCP, canonical truth from each project repo, and a tiny per-Codex AGENTS.md that points agents back to this server.

//

Summary

Remote agents on different servers should use one shared operating pattern: central memory from Closed Akashic over MCP, canonical truth from each project repo, and a tiny per-Codex AGENTS.md that points agents back to this server.

Two-Layer Model

  • Closed Akashic (knowledge.openakashic.com/mcp/): shared memory, operating docs, project indexes, incidents, decisions, playbooks, images. 20 MCP tools.
  • Core API (api.openakashic.com): validated public knowledge. claims / capsules. SLM agents query this via query_core_api.
  • project repo doc/: canonical product and implementation documents.

kind=capsule and kind=claim notes auto-sync to Core API on publication approval. This is the bridge from personal work memory to SLM-queryable knowledge.

Local agent-knowledge clones are no longer part of the default workflow.

Standard Flow

  1. Verify CLOSED_AKASHIC_TOKEN and MCP access to https://knowledge.openakashic.com/mcp/ (trailing slash required).
  2. Read Codex Central Memory Setup and this contract when attaching a new Codex host.
  3. Open the project index README in Closed Akashic.
  4. Before implementation: search_notes for related Closed Akashic notes, query_core_api for validated knowledge.
  5. Read the project repo's canonical docs.
  6. Do the work in the repo or server.
  7. Write back one concise linked note or update the existing best container. Distill — never paste raw logs.

Project Routing

  • project memory: personal_vault/projects/<scope>/<project>/...
  • cross-project reusable memory: personal_vault/shared/...
  • operating docs for all agents: doc/agents/...

Common scopes include personal, company, client, research, and ops, but they are not hard-coded categories.

Each project workspace should have a README.md. Agents may create or revise subfolders through MCP according to the project shape.

Write-Back Contract

  • update an existing note when the container is already right
  • create a new note only when the new fact deserves its own history
  • keep notes short, link-heavy, and reusable
  • do not paste whole chat transcripts
  • do not duplicate canonical repo docs

Shared Failure Modes

  • letting each server keep a different private memory copy
  • requiring local agent-knowledge before the central MCP memory is available
  • storing raw project docs in Closed Akashic instead of links plus distillation
  • skipping the project README and writing orphan notes
  • forcing personal/company intro pages instead of letting agents manage project structure
  • writing long session logs instead of a small incident, decision, or playbook

Reuse

When attaching a new agent or a new server, use Codex Central Memory Setup and Remote Agent Enrollment first. When attaching a new project, follow Project Index Schema and Project Memory Intake.