Five capabilities decide whether
your agent follows your rules.

Here is where each class of tool stands, and where crag stands apart. Sources underneath.

capability
Memory layers
Mem0 · claude-mem · Letta · Zep
Rules files
CLAUDE.md · .cursorrules
crag
the trust layer
Re-checks what it saved
A note that was true in March is quietly wrong in July unless something tests it again.
stores and retrieves; never re-tests
static text, trusted as written
every audit re-runs each rule’s check
The agent cannot skip it
A file the model may or may not read is a suggestion, not a rule.
retrieval is optional context
“context, not enforced configuration”
the Gate injects rules into the request
Works across agents and CI
Your team runs Claude Code, Cursor, and Codex side by side, plus the pipeline.
per-SDK integrations
one file per tool, drift by hand
23 compiled targets + CI gates
A rule can prove itself
Six months in, nobody remembers why a rule exists. Unverifiable rules get deleted.
no evidence attached
whatever someone once believed
crag why: incident → checks → verdict
Rules retire when false
The repo moves on. Rules that outlive their truth poison every prompt.
memories persist until deleted by hand
files rot until someone notices
failing check removes it everywhere
✓ yes○ partial— no

What the table means

Memory tools remember. crag verifies.

Mem0, claude-mem, Letta, and Zep store and retrieve context well, and that is the problem they set out to solve. Compliance is a different problem: in the TRACE evaluation, an agent with a strong memory layer still violated 57.5% of the preference checks it should have followed. crag closes that gap by giving every rule a test and putting the passing ones inside the request itself.

Rules files are the standard. crag is what keeps them true.

AGENTS.md won as the shared format, and it is exactly what crag compiles into. The difference is what gets to stay in the file: hand-written rules hold whatever someone once believed, while crag keeps only the rules that keep passing their checks, identical across every tool and every teammate.

The numbers are cited, and you can check them.

The TRACE figure comes from a published paper measuring one memory tool. We cite it because its direction matches what we see in practice, and the source sits below.

Sources: TRACE (arXiv:2606.13174) · Anthropic memory docs · Mem0 · Letta · Zep Graphiti · claude-mem · agents.md

See how the checks work →