Many agents need the same rules. This is what comes next.
Today, Mneme checks one local repository. For many agents, every agent needs the exact same rule, checked the same way, with a clear record of what happened. This page shows that target without calling it a shipped demo.
Keep every agent on the same rules.
What works today is marked clearly. The shared pieces are the work still ahead.
Retrieval remains deterministic. No agent may introduce sentence-transformers or a vector database unless the recorded decision is explicitly amended.
-
01 Decision
A version-controlled repository records the deterministic retrieval boundary.
mneme_retrieval_deterministic · decision corpus v17Current -
02 Agent change
Another agent proposes a retrieval rewrite that conflicts with the rule.
actor B · add vector databaseIllustrative input -
03 Shared gateTARGET
Every agent would need the same policy version at the same enforcement boundary.
TARGET · shared policy + actor identityRequires Layer 2 -
04 Shared traceTARGET
A blocked change and its corrected retry must remain connected across actors.
TARGET · persistent cross-agent evidenceRequires Layer 2
What is demonstrable today
Layer 1 proves the local enforcement primitive
Current: a local repository stores the decisions, retrieves applicable constraints deterministically, and blocks a violating task before one agent executes it.
CURRENT version-controlled project decision corpus
CURRENT deterministic retrieval and pre-flight enforcement
CURRENT local-repository decision trace
TARGET shared policy distribution across actors
TARGET actor + decision-version identity
TARGET persistent deny → correction → allow traceWhat a real multi-agent proof must add
Each result must say exactly which set of project rules the agent received.
Parallel agents cannot skip the rules by using a different tool or session.
The record must connect the denied attempt, the correction, and the allowed result across agents.
Common questions
Is this a runnable multi-agent demo?
No. It is a target contract. The stages that need shared rules and persistent cross-agent records are marked as work still to do.
What part works today?
One repository can store clear architecture rules, find the relevant ones, and block a violating task before a local agent runs it.
What would make this real?
All agents need the same rule version, the same check point, and one record that connects a block to its compliant retry.
Start with one governed repository.
Start with a governed repository today. The shared, multi-agent layer is the next capability to prove.
pipx install "mneme-hq>=0.5.1"