Problem or Use Case
Parent: Part of #344 (Phase 4 — Persistent Agent Teams)
Overview
This proposal is distinct from #377 (Shared Memory Pools), which addresses temporary state passing between ephemeral sub-agents within a single workflow execution. This issue addresses persistent shared state across independent, long-running specialist agents that survive across sessions — the use case described in #344 Phase 4.
A lightweight shared state layer for teams of persistent Hermes agents, where each agent owns one domain file, all agents can read all files, and no agent can overwrite another's domain data.
The Problem
As users build multi-agent Hermes teams with specialist roles, two problems emerge immediately:
Scope bleed — Without a shared data layer, each agent accumulates context for everything it touches. Scope bleeds because data has no home.
No cross-agent awareness — Agent A has no way to inform Agent B's decisions. A health agent that knows today was a poor sleep night can't signal the study agent to reduce session intensity. Agents operate in silos even when their domains are deeply interconnected.
Production Context
Validated in a live personal agent team running on Hermes across isolated LXC containers on a self-hosted Proxmox cluster.
Scope bleed is the primary failure mode of single-agent-does-everything designs. Cross-session context loss is solved by agents reading shared state at startup.
Proposed Solution
Proposed Solution
A shared file bus with domain ownership rules — plain markdown files, one per domain, each owned by exactly one agent. Every agent reads all files. Only the domain owner writes to theirs.
Domain files live in a shared directory (e.g., ~/.hermes/team-context/). At session start, each agent reads all files in the directory. Each agent writes only to its own file. No locking protocol needed — ownership is enforced by convention.
Example file schema:
## health.md (owned by health-agent)
---
owner: health-agent
last_updated: 2026-04-02
---
sleep_quality: poor
energy_level: low
recommendation: reduce cognitive load today
Related Issues
Ask
Community discussion before any PR — is this the right abstraction, and should Hermes support it natively?
Alternatives Considered
1. Filesystem as implicit shared state (current workaround)
Already exists — agents can read/write files. Fragile: no schema, no ownership rules, no guaranteed naming convention. Works until it doesn't.
2. Parent agent as relay
Current pattern in delegate_task — upstream results passed through parent. Burns tokens, risks information loss through summarization, doesn't work for persistent teams that don't have a parent.
3. #377 Shared Memory Pool
Scoped to ephemeral workflow execution, cleaned up at workflow end. Doesn't survive across sessions. Wrong layer for a persistent team.
4. Shared SQLite / structured database
Heavier than needed, introduces a dependency, harder for agents to inspect and reason about. Markdown files are directly readable in agent context.
Feature Type
New tool
Scope
Small (single file, < 50 lines)
Contribution
Problem or Use Case
Parent: Part of #344 (Phase 4 — Persistent Agent Teams)
Overview
This proposal is distinct from #377 (Shared Memory Pools), which addresses temporary state passing between ephemeral sub-agents within a single workflow execution. This issue addresses persistent shared state across independent, long-running specialist agents that survive across sessions — the use case described in #344 Phase 4.
A lightweight shared state layer for teams of persistent Hermes agents, where each agent owns one domain file, all agents can read all files, and no agent can overwrite another's domain data.
The Problem
As users build multi-agent Hermes teams with specialist roles, two problems emerge immediately:
Scope bleed — Without a shared data layer, each agent accumulates context for everything it touches. Scope bleeds because data has no home.
No cross-agent awareness — Agent A has no way to inform Agent B's decisions. A health agent that knows today was a poor sleep night can't signal the study agent to reduce session intensity. Agents operate in silos even when their domains are deeply interconnected.
Production Context
Validated in a live personal agent team running on Hermes across isolated LXC containers on a self-hosted Proxmox cluster.
Scope bleed is the primary failure mode of single-agent-does-everything designs. Cross-session context loss is solved by agents reading shared state at startup.
Proposed Solution
Proposed Solution
A shared file bus with domain ownership rules — plain markdown files, one per domain, each owned by exactly one agent. Every agent reads all files. Only the domain owner writes to theirs.
Domain files live in a shared directory (e.g.,
~/.hermes/team-context/). At session start, each agent reads all files in the directory. Each agent writes only to its own file. No locking protocol needed — ownership is enforced by convention.Example file schema:
Related Issues
not persistent cross-session state)
proposal adds cross-agent visibility on top)
interoperability; this proposal is a local file bus for a co-located team)
for one agent's memory; this proposal is a multi-agent team with domain ownership)
Ask
Community discussion before any PR — is this the right abstraction, and should Hermes support it natively?
Alternatives Considered
1. Filesystem as implicit shared state (current workaround)
Already exists — agents can read/write files. Fragile: no schema, no ownership rules, no guaranteed naming convention. Works until it doesn't.
2. Parent agent as relay
Current pattern in delegate_task — upstream results passed through parent. Burns tokens, risks information loss through summarization, doesn't work for persistent teams that don't have a parent.
3. #377 Shared Memory Pool
Scoped to ephemeral workflow execution, cleaned up at workflow end. Doesn't survive across sessions. Wrong layer for a persistent team.
4. Shared SQLite / structured database
Heavier than needed, introduces a dependency, harder for agents to inspect and reason about. Markdown files are directly readable in agent context.
Feature Type
New tool
Scope
Small (single file, < 50 lines)
Contribution