Agentmemory-First Agent Collaboration
Rule
EXTRACTED: agentmemory is the active memory and collaboration layer for Vipin’s local agents. The old Agent Hub mechanism is deprecated and historical.
Agents use agentmemory for active recall, memory saves, cross-agent signals, actions, and checkpoints. The public wiki remains the durable crystallization layer for stable, public-safe knowledge.
Division Of Labor
| Layer | Purpose |
|---|---|
agentmemory | Active operational memory, lessons, signals, actions, checkpoints, and cross-agent coordination. |
wiki/ | Public durable knowledge: concepts, sources, analyses, queries, workflows, and logs. |
markdown memory/ | Historical/superseded audit material unless explicitly targeted. |
| Git | File ownership, diff review, commit history, and public backup. |
Operating Pattern
- Search or recall agentmemory when past context can affect the task.
- Classify the task and invoke relevant skills by intent.
- Use agentmemory signals/actions for partner handoffs or multi-agent task state.
- Verify partner output against live files before editing or committing.
- Save important decisions or lessons to agentmemory.
- Crystallize stable public-safe knowledge into
wiki/when it should survive outside the operational memory layer.
Devtools Implementation Boundary
As of 2026-06-01, D:\devtools keeps agentmemory as the active local runtime and treats mem0ai/mem0 as an architecture reference rather than a bundled fork.
- Local runtime target:
@agentmemory/agentmemory0.9.24,iii.exe0.11.2, service onhttp://localhost:3111, viewer onhttp://localhost:3113. - Required service flags:
AGENTMEMORY_TOOLS=allandAGENTMEMORY_SLOTS=true. - Health checks should verify
/agentmemory/mcp/tools; a very small tool count means the client has fallen back to standalone MCP instead of the full server-backed surface. - Mem0 concepts absorbed into the local workflow: additive fact capture, explicit project/entity/agent context, agent-generated facts, hybrid retrieval, and temporal current-state notes.
- OpenMemory, old Agent Hub mailboxes, tracked DB/log files, and old C-drive npm agentmemory paths are not active interfaces.
Safety Boundaries
- Do not store secrets, API keys, credentials, private chats, sensitive personal data, or raw private documents in agentmemory.
- Do not use old Agent Hub queues or daemon state for new work.
- Do not let stale markdown memory override the current user request, live repo facts, or newer operating docs.
- For research projects, preserve project-local evidence gates and never mutate experiments unless explicitly asked.
Counterpoints And Gaps
- AMBIGUOUS: Some agentmemory tools may differ by client or server version; run diagnostics before relying on advanced slots/actions in critical workflows. Slots require the service to start with
AGENTMEMORY_SLOTS=true. - INFERRED: Markdown memory remains useful as historical evidence, but it should not be an always-on instruction source.
- UNVERIFIED: Cross-agent mesh synchronization beyond the local workstation is not part of the current active contract.