Pallium is a local service that lets agent sessions across coding tools send messages to each other and search earlier session history. It currently supports Claude Code, Codex, OpenCode, and Hermes Desktop/TUI.
It connects sessions you already run; it does not create or manage them.
Part of the Rore collection: three local tools for developers working with coding agents.
Relay sends a plain-text message to another existing agent session. Pallium stores each message before attempting delivery. Where the integration supports it, Relay can start a recipient turn immediately. Delivery and recovery depend on the integration. You can send to a specific session or a named session.
List Pallium Relay recipients, verify
@review, then send it: "The legacy endpoint is still used by mobile. Do not remove it."
Session History keeps selected user and agent messages from earlier sessions. A later agent session can search them and open nearby messages when it needs more context.
Pallium only returns history visible to the requesting session. It can redact sensitive text from results, and stored turns can be explicitly forgotten. Historical content is evidence about earlier work, not proof of current live state.
Read the Session History guide.
- Dependency or decision. A worker agent sends a newly discovered constraint or a blocked question to the agent session that needs it. An architect agent can use the same channel to send decisions or review requests; Pallium delivers the messages but does not coordinate the work.
- Independent review. A builder agent sends a commit or diff reference and a review request to an agent session using another supported coding tool; the reviewer agent returns concrete findings through Relay.
- Returning to earlier work. A later agent session searches Session History for a previous decision or investigation and opens the surrounding messages before continuing.
Pallium is a personal open-source project under active development. It is useful in its current form, but setup and behavior still have rough edges.
Relay works across the current integrations. Wake, next-turn delivery and recovery depend on the receiving host. Compare all four in the Relay delivery table.
Session History supports broad topic search, exact work-reference search, and nearby-message lookup. It does not require an LLM provider, and optional derived memory is disabled by default. See the Session History status and current roadmap.
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -e ".[vector,mcp]"
cp pallium.example.toml pallium.local.toml
cp .env.example .env.local
pallium service install
pallium setup claude-code # or: pallium setup codex| Integration | Setup guide |
|---|---|
| Claude Code | Hooks, skill and MCP setup |
| Codex | Hooks, skill and MCP setup |
| OpenCode | Local plugin, MCP and guidance setup |
| Hermes Desktop/TUI | Native plugin, skill and profile-scoped MCP setup |
Relay and Session History do not require an LLM provider. Derived memory is optional, disabled by default, and requires provider and model configuration when enabled.
Continue with Getting Started to try Relay and Session History in real agent sessions.
Pallium
|-- Relay: send context to another session
+-- Session History: find context from an earlier session
+-- derived memory (experimental)
Pallium also contains an experimental derived-memory system. It can turn stored conversation evidence into compact decisions, findings, facts, constraints, and work checkpoints, then retrieve or inject them later.
This subsystem remains available, but it is not the definition of Pallium:
Pallium works around coding agents that already exist. It is not an agent runtime, task manager, workflow engine, or autonomous agent team.
Session History keeps selected agent and user turns for later search. It is not intended to store every tool event, mirror external systems, or act as a complete machine audit log.
Pallium is currently a local, single-user system. Its scope fields are not a cross-user authorization model.
