Skip to main content

Memory

PilotDeck includes a PilotDeck Memory Core long-term memory system. It captures important knowledge from conversations, retrieves relevant context in new sessions, and injects it into prompts so your agent can remember the WorkSpace.

Core Capabilities​

Automatic Capture

Extract important information after each conversation turn and store it as memory

Smart Retrieval

Retrieve relevant context when a new session or turn starts

WorkSpace-Aware

Rank memory with the current WorkSpace context in mind

Cross-Session

Carry knowledge from earlier sessions into future conversations

How It Works​

Memory has two stages: Capture and Retrieve.

Turn Completedconversation transcriptCaptureextract key factstag session / timestampPilotDeck Memorypersistent knowledge store~/.pilotdeck/memoryNew Turnlatest query + contextRetrievesearch relevant memoryrank by query and workspaceSystem Contextinject retrieved factsbefore model requestread pathwrite path

Capture​

After each turn, Memory:

  1. Converts conversation messages into standardized ContextMemoryMessage values
  2. Calls captureTurn in PilotDeck Memory to extract knowledge
  3. Stores the result with sessionKey and timestamp metadata

Retrieve​

At the start of a new session or turn, Memory:

  1. Builds a query from the latest user message and recent context
  2. Calls retrieveContext in PilotDeck Memory
  3. Injects the retrieved result as systemContext

Retrieval Modes​

The system decides whether retrieval is needed:

memory:
retrievalMode: auto

Configuration​

memory:
enabled: true
rootDir: ~/.pilotdeck/memory # optional, defaults to PilotHome/memory
retrievalMode: auto # auto | explicit

Memory data is stored in ~/.pilotdeck/memory by default.

Integration with Context​

Memory is part of PilotDeck's Context system. Context manages the full prompt context for each request:

ComponentResponsibility
PromptAssemblerBuild the System Prompt from instructions, skills, memory, and tools
MessageProjectorProject conversation history and control context size
TokenBudgetManagerTrack token budget and avoid context overflow
CompactionEngineCompress old messages when context grows too long
MemoryResolverRetrieve long-term memory from PilotDeck Memory and inject it into context

Memory Position in Context​

System Prompt
├── Core instructions
├── Project instructions
├── Skills context
├── Memory context ← injected by MemoryResolver
└── Tool descriptions

Attachment Mechanism​

Memory retrieval results are converted by MemoryAttachmentBuilder into the same attachment structure used by files, images, and other context attachments:

{
systemContext: "Based on previous discussions, this project uses React 18 + TypeScript...",
diagnostics: [],
metadata: { trace: ..., debug: ... }
}

Diagnostics​

Diagnostic codeDescription
memory_context_emptyNo relevant memory found, info level
memory_provider_errorPilotDeck Memory service error, warning level
提示 · Tip

If Memory quality is poor, inspect ~/.pilotdeck/memory and verify capture is working.