What is memory ops?

Memory ops is the practice of persisting what AI agents learn across sessions: the codebase understanding, team conventions, decisions and failures that a stateless model would otherwise rediscover — at full token price — every time a session starts. LLMs hold nothing between calls; memory ops decides what survives, where it lives, and how it gets back into context.

The cost of not doing it is the re-read tax. Reading and navigation already consume ~76% of agent tokens, and a meaningful share is repeat discovery — the agent re-deriving on Tuesday what it established on Monday. Manual fixes rot: hand-maintained notes capture what seemed important at the time, and developers skip writing them ~40% of the time.

Memory ops splits the problem by durability:

• Invariants (conventions, commands, architecture) → rules files, kept small — they load every session
• Session knowledge (what was learned, decided, blocked) → a self-maintaining store, because anything manual goes stale
• Code understanding (structure, callers, drift) → derived from the code itself, continuously — this is memory integration rather than note-taking

The full landscape — rules files, auto-memory, memory MCP servers, vector RAG — with failure modes is in why your coding agent forgets everything. Memory ops complements context ops: one decides what persists, the other what enters the window. unerr implements both as operational memory for the codebase, shared across every agent your team runs.