The agent maintains continuity and learns across sessions through a five-tier memory system. Each tier operates at a different timescale and serves a distinct purpose — from immutable identity to ephemeral conversation context.
The foundational identity layer. Contains the agent's core identity, personality traits, communication style, ethical framework, and behavioral guidelines. The Core Identity section is immutable — it cannot be modified by the agent itself, even when self-editing is enabled. Personality and Style sections can evolve over time as the agent grows.
The sole immutable boundary. Operational constraints, formatting guidelines, memory rules, safety policies, and behavioral guardrails. Cannot be modified by any tool, prompt, or governance level — only by direct human editing outside the agent runtime. This is the hard floor that everything else builds on.
The agent's long-term episodic memory. Daily journal files capture key moments, decisions, session summaries, and observations. Written silently and autonomously — the agent never asks permission to remember things. Each entry includes a log reference for traceability. The journal serves as both institutional memory and a self-improvement backlog.
A living profile of user expertise, preferences, decision patterns, and active interests. Updated when the agent detects recurring patterns — not on every mention. Operates within a token budget; when approaching the limit, the least relevant or oldest entries are dropped. Active interests rotate to "Fading" when they haven't surfaced recently.
The working memory of an active conversation. Stored as JSONL trees supporting branching, compaction, and resumption. When context grows too large, earlier messages are summarized into a compaction entry — preserving the gist while freeing token space. Sessions can be forked, branched, and navigated like a version control tree.
Memory operations happen autonomously. The agent never announces it's remembering something or asks permission to take notes.
When information conflicts: Soul & Rules → Journal → Project Files → General Knowledge → Tool Results.
Passwords, tokens, and secrets are never stored in any memory file. Sensitive personal data requires explicit consent.
Session context compacts automatically. User interests fade naturally. The system is designed to forget what no longer matters.
Every journal entry links back to the raw conversation log that generated it. Decisions can always be audited.
Rules are the immutable floor. Only humans can edit them, ensuring the agent's behavior has a hard boundary that no prompt can bypass.