Memory
Agents forget everything between turns unless you give them memory. Koryphaios has four layers, each with a different scope, plus explicit rules and preferences the critic enforces.
The layers
Universal memory
Facts true across every project — who you are, your stack preferences, your conventions. Applied everywhere.
Project memory
Scoped to the current project (or workspace root). Architecture notes, gotchas, decisions specific to this codebase.
Session memory
Per-conversation persistent storage — carries context within a session without polluting the global stores.
Rules & preferences
rules.md is a hard contract the critic enforces on every change; preferences.md captures softer “how I like work done” guidance. Rules are always applied.
Workspace-shared memory
Open a folder of projects as a workspace and they share one memory/rules store at the root, instead of each project reinventing it. A sidebar toggle switches between the project scope and the whole workspace.
How memory reaches the agent
Enabled memory is injected into the agent's context automatically, but under a token budget — Koryphaios estimates the cost and adds memory greedily up to the limit rather than blindly stuffing the window. You control which layers are on and the max context tokens in Settings → Memory. Agents can also be allowed to write memory (e.g. record a decision during compaction), gated by a single toggle that's the same setting shown in both the Memory and Agent tabs.
Memory vs. notes
Memory is prose the agent always considers; notes are a linked knowledge graph the agent queries on demand and can pin into context. Use memory for “always remember this,” and notes for a browsable body of knowledge. See the Notes page.