Memory Plane
Memory is one of the hardest problems in agentic systems, and one of the most poorly solved. The chatbot model treats memory as a growing list of messages that eventually overflows. The Agentic OS treats memory as a managed, tiered, disciplined resource.
The Problem
Language models have a fixed context window. Everything the model can “think about” must fit in that window. Without memory discipline, agentic systems face a brutal tradeoff: either keep everything (and overflow) or discard things (and lose critical context).
Operating systems solved this problem decades ago with virtual memory, paging, caching, and tiered storage. The Agentic OS applies the same principles to cognitive context.
Memory Tiers
block-beta
columns 1
WM["Working Memory\n(Hot — active context, ephemeral)"]
EM["Episodic Memory\n(Warm — interaction histories, decision records)"]
SM["Semantic Memory\n(Cool — domain knowledge, learned patterns)"]
OS["Operational State\n(System metadata, process states, configs)"]
AH["Audit History\n(Cold — immutable action log, for humans)"]
style WM fill:#c0392b,stroke:#e74c3c,color:#fff
style EM fill:#d35400,stroke:#e67e22,color:#fff
style SM fill:#2980b9,stroke:#3498db,color:#fff
style OS fill:#27ae60,stroke:#2ecc71,color:#fff
style AH fill:#2c3e50,stroke:#34495e,color:#e8f0fe
Working Memory
The immediate context of the current task. This is what the active process can “see” right now. It is small, focused, and ephemeral.
- Current task definition
- Relevant retrieved context
- Intermediate reasoning results
- Active plan state
Episodic Memory
Records of what has happened. Structured summaries of past interactions, decisions, and outcomes. Episodic memory answers: “What did we do before?”
- Compressed interaction histories
- Decision records
- Outcome summaries
- Failure logs
Semantic Memory
Long-term knowledge that is not tied to specific interactions. Facts, concepts, domain knowledge, learned patterns. Semantic memory answers: “What do we know?”
- Domain knowledge bases
- Learned patterns and heuristics
- Entity relationships
- Organizational knowledge
Operational State
System-level metadata about the current state of the Agentic OS itself. Not task content, but system health and status.
- Active processes and their states
- Resource utilization
- Policy configurations
- Pending approvals
Audit History
An immutable record of every action, decision, and policy evaluation. This is not for the model to reason over — it is for humans to inspect, debug, and verify.
Memory Operations
| Operation | Purpose |
|---|---|
| Store | Write information to the appropriate tier |
| Retrieve | Pull relevant information into working memory |
| Compress | Summarize detailed records into compact representations |
| Evict | Remove information that is no longer needed |
| Reconcile | Resolve contradictions between memory tiers |
| Prune | Remove outdated or contradicted information |
Memory Discipline
The key insight is that memory must be managed, not just accumulated. This means:
- Selective retrieval — Only pull into working memory what the current task needs
- Strategic compression — Summarize rather than discard; preserve the signal, reduce the noise
- Contradiction detection — When new information conflicts with stored memory, resolve it explicitly
- Budget enforcement — Each process has a context budget; the memory plane ensures it is respected
The OS Parallel
Just as an OS pages memory to disk when RAM is full and pages it back when needed, the memory plane moves context between tiers:
- Hot context lives in working memory (fast, small)
- Warm context lives in episodic memory (summarized, retrievable)
- Cold context lives in semantic memory (indexed, searchable)
- Frozen context lives in audit history (immutable, archival)
This tiering is what makes agentic systems efficient. Without it, context windows overflow and performance collapses.