Forge Memory Guide

Forge memory now lives in the staged store under ~/.fluid/store/. The store is shared by forge data-model runs, MCP tools, provider scorecards, and audit/history capture.

CLI entry point

fluid memory status
fluid memory show project
fluid memory show semantic
fluid memory search semantic "customer order model"
fluid memory clear --ns memory/semantic

The command group is:

fluid memory <status|show|save|clear|search>

Store layout

~/.fluid/store/
├── llm/
├── memory/
│   ├── project/
│   ├── team/
│   ├── episodic/
│   └── semantic/
├── discovery/
├── skills/
├── history/
└── audit/
NamespaceWhat it stores
memory/projectProject-scoped preferences and summaries
memory/teamShared team conventions when configured
memory/episodicTime-ordered forge episodes
memory/semanticSimilarity-searchable forged model summaries
historyVersioned artifact snapshots from write tools
auditCatalog reads, MCP mutations, and forge events

Semantic memory is opt-in:

FLUID_COPILOT_SEMANTIC_MEMORY=1 fluid forge data-model from-intent intent.yaml -o out.fluid.yaml

What memory does

When enabled, memory helps Forge remember stable preferences such as:

  • project naming conventions
  • preferred modeling technique
  • prior source/catalog scopes
  • recurring entity vocabulary
  • prior model patterns that can improve later drafts

Memory remains advisory. Explicit CLI flags, current user input, current catalog/DDL evidence, and validation gates win over saved memory.

Privacy and credentials

Memory is not a raw session dump. It should not contain:

  • API keys
  • tokens
  • raw sample rows
  • full source data extracts
  • private keys

Catalog credentials live in the OS keyring and ~/.fluid/sources.yaml references; MCP source-catalog calls pass credential ids, not raw secrets.

Legacy memory

Older workspaces may still contain:

.fluid/copilot-memory.json

The current CLI can read that legacy file with a one-time notice, but new writes land in ~/.fluid/store/memory/project/.