Cross-agent memory
Ironspire's memory system (powered by the Corpus engine) gives your agents persistent, shared knowledge. Agents automatically capture observations as they work, and those observations are recalled in future sessions to provide continuity.
Memory is a Forge feature. Starter-tier users do not have access to the memory system.
How it works
- Agents observe: as an agent works, it captures notable findings as memories
- Memories persist: stored in a local SQLite database with full-text search
- Context injection: before each message, relevant memories are recalled and injected into the agent's prompt as context
- Decay over time: old or stale memories gradually lose relevance, keeping context fresh
This all happens automatically. You do not need to configure anything for memory to work.
Memory types
The system classifies every memory as one of three types:
| Type | Description | Example |
|---|---|---|
| Fact | Stateless knowledge or information | "Project uses TypeScript with strict mode" |
| Event | A timestamped occurrence or action | "Agent deployed v2.3 to staging on March 1" |
| Pattern | A recurring behavior or workflow insight | "Agent prefers JSON over YAML for configs" |
Scopes
Memories have a privacy scope that controls who can access them:
- Private: only the agent that created it can recall it
- Project: shared with all agents in the same project
- Global: accessible to all agents across all projects
Confidence levels
Each memory carries a confidence indicator:
- Low: uncertain, needs verification
- Medium: likely accurate
- High: confirmed and reliable
- Critical: essential knowledge, should not decay
The Memory panel
Open the Memory panel from the left sidebar (under Panels). It has three tabs:
Memories tab
Browse, search, and filter all stored memories.
- Search: full-text search across titles and content
- Type filter: toggle Fact, Event, and Pattern
- Scope filter: toggle Private, Project, and Global
- Tag filter: filter by custom tags
- Agent filter: show memories from a specific agent
Each memory card shows the title, type badge, scope badge, confidence indicator, creating agent, and timestamp. Click to expand and see the full content, related memories, and lifecycle state.
Summaries tab
View automatically generated summaries at multiple levels:
- Session: created when a conversation completes
- Daily: generated at end of day, captures key achievements
- Milestone: generated after 7+ daily summaries accumulate
- Project: high-level project outcomes
Summaries are grouped by date with friendly labels ("Today", "Yesterday", etc.).
Conflicts tab
When agents record contradictory information, conflicts appear here. Each conflict card shows the two conflicting memories side by side with resolution options:
- Keep A or Keep B: choose one
- Merge: combine both into a new memory
- Dismiss: ignore the conflict
How agents read memories
Before each agent message, Ironspire recalls relevant memories and injects them as context, organised into three layers:
- Patterns: behavioural insights and workflow preferences
- Recent events: recent actions and outcomes
- Relevant facts: knowledge base entries matching the current context
If no memories match the current context, nothing is injected and there is no overhead.
You do not need to tell agents to "remember" things. Memory capture and recall happen automatically. You can review and manage memories in the Memory panel at any time.
Background automation
The memory system runs two background processes:
- Reflection: periodically analyses agent behaviour and creates pattern memories
- Decay: gradually reduces the relevance weight of old memories
The Memory panel shows the status of both processes at the bottom.