What is persistent memory for AI coding agents?
Persistent memory for an AI coding agent is a record of decisions, failed approaches, and rules that survives past a single session — so the agent doesn't re-investigate settled questions every time you start a new chat. LinkLore implements this as an AI Native Note: typed entries an agent reads at the start of a session and writes to as it works.
Why it matters
An AI agent's context window resets when a session ends. Without persistent memory, every new chat starts from zero — the agent can re-read your code, but not the decisions behind it, the approaches you already ruled out, or the rules your team settled last week.
How this differs from CLAUDE.md and built-in memory
CLAUDE.md and similar instruction files persist on disk, but they don't know when they're wrong — a stale line sits there looking as confident as a current one. Claude Code's built-in auto memory is machine-local and isn't shared across teammates or other tools. See the full comparison →
What it looks like in practice
> The walk-up behavior in local() looks like a bug — preparing a fix. ⚑ lore surfaced · lr-22c2c510 "local() walk-up is by design — confirmed decision. Rationale: parent-dir resolution is the contracted behavior." > Reverting. The code was correct — my fix would have broken it.
A real session. The agent was about to "fix" code that was actually correct — a three-week-old decision note, surfaced automatically, stopped it.
The evidence
Recalling a settled decision costs about 2,000 tokens; re-deriving it in a fresh session cost hundreds of thousands. From an 11-month study across 26 real projects (5,699 commits). Read the study ↗
When you don't need this
If you work alone on something that fits in a single session, or rarely revisit old decisions, plain markdown is enough. Persistent memory pays off once sessions span days or weeks, more than one agent touches the project, or "why did we do it this way" comes up more than once.