Concepts

Memory engineering vs related layers

Discipline	Question
Context engineering	What fits in this prompt?
Harness engineering	How does one run execute safely?
Memory engineering	What is true across runs, and who may write it?
Loop engineering	What system keeps re-prompting and verifying?
Fleet engineering	How do many agents share governance?

Memory is not “more context.” Memory is state with policy: tiers, gates, budgets, hygiene.

The four tiers

1. Scratch — session working memory; high churn, low trust
2. Episodic — narrative of what happened; medium trust
3. Durable facts — small high-trust invariants; gated writes
4. Retrieved — computed pack for one inference under budget

Confidence tags

Tag	Meaning	Can be durable?
observed	Agent saw evidence	After verify
decided	Human/process decision	Yes
hypothesis	Guess	Never

Promotion

scratch  --(session end)-->  episodic  --(verifier/human)-->  durable

Skipping gates is an anti-pattern (see anti-patterns.md).

Stores

Default: markdown files in git (MEMORY.md, MEMORY-STATE.md).

Optional later: vector DB, Mem0, team wiki — behind the same tier policy. Patterns stay store-agnostic.

Relationship to STATE.md (loops)

File	Job
STATE.md	What the loop is working on now
MEMORY-STATE.md	What the project/agent should remember across time

Do not merge them into one infinite document.