What happens when
your AI agent remembers

Five sessions. Same task. One agent starts from zero every time. The other compounds. Watch the difference.

↓ scroll to begin
1
Session 1

Add user authentication

First session. Both agents start from zero. No difference yet.

without tenet
$ claude "add user auth with JWT"
Creating src/auth.ts...
Added bcrypt for password hashing
Added JWT token generation
✓ Auth module created
Session ends. Context gone.
with tenet
$ tenet "add user auth with JWT"
Creating src/auth.ts...
Added bcrypt for password hashing
Added JWT token generation
✓ Auth module created
3 memories saved. 1 training tuple.

Context accumulated: 3 memories

decision JWT chosen over session cookies — stateless API
feature src/auth.ts — bcrypt hashing, token generation, middleware
discovery bcrypt rounds=10 causes 200ms latency on login
2
Session 2 — next day

Add password reset flow

The stateless agent doesn't know auth exists. It builds a second auth system. TENET extends what's already there.

without tenet
$ claude "add password reset"
Creating src/password-reset.ts...
Adding bcrypt for hashing...
⚠ Duplicate: bcrypt already in auth.ts
Creating new JWT helper...
⚠ Duplicate: JWT already in auth.ts
Different token format than auth.ts
✗ Two auth systems, incompatible tokens
with tenet
$ tenet "add password reset"
Loading 3 memories about auth...
Found existing auth in src/auth.ts
Reusing JWT helper + bcrypt config
Adding resetToken to existing schema
✓ Password reset extends auth module
✓ Same token format, same bcrypt config

Context accumulated: 6 memories

decision Reset tokens expire in 1 hour, stored in same JWT format
feature src/auth.ts extended — resetPassword(), verifyResetToken()
discovery Email service needs SMTP config — deferred to session 3
3
Overnight — while you sleep

Agents run autonomously

TENET agents work overnight. They propose changes, eval measures the delta, good changes are kept. Your codebase improves while you sleep.

tenet overnight
02:14 AM
agent:test-coverage Round 1: added 12 tests for auth module
coverage: 14.2% → 18.7% +4.5% kept ✓
agent:test-coverage Round 2: added edge case tests
coverage: 18.7% → 21.3% +2.6% kept ✓
agent:code-quality Round 1: extracted auth middleware
quality: 0.48 → 0.52 +0.04 kept ✓
03:47 AM — PR #14 created
Training buffer: 3 new tuples captured

Context accumulated: 11 memories + 3 training tuples

training {state: auth untested, action: add JWT tests, reward: +4.5%}
training {state: edge cases missing, action: add boundary tests, reward: +2.6%}
training {state: inline middleware, action: extract to module, reward: +0.04}
4
Session 4 — one week later

Add OAuth (Google, GitHub)

A week of sessions. The stateless agent has no idea what auth strategy you chose. TENET has 11 memories and a trained policy.

without tenet
$ claude "add Google OAuth"
Installing passport-google-oauth20...
Creating new session system...
⚠ Conflicts with existing JWT auth
Creating src/oauth.ts...
⚠ Different user model than auth.ts
✗ OAuth and JWT auth are disconnected
✗ No tests (doesn't know you have 21%)
with tenet
$ tenet "add Google OAuth"
Loading 11 memories... policy head consulted
Decision: JWT auth (session 1) → extend, don't replace
Extending src/auth.ts with OAuth provider
Same user model, JWT issued after OAuth
✓ OAuth unified with existing JWT flow
Adding tests (coverage currently 21.3%)
✓ 6 new tests, coverage → 24.1%
5
Session 5 — debugging

"Why is login slow?"

The question that separates memory from no memory. The stateless agent can only guess. TENET has the history.

without tenet
$ claude "why is login slow?"
Possible causes:
- Database query optimization
- Network latency
- Session creation overhead
Generic advice. Doesn't know your stack.
Doesn't know bcrypt is the bottleneck.
Doesn't know you chose JWT over sessions.
with tenet
$ tenet "why is login slow?"
Searching 14 memories...
Found: session 1 — bcrypt rounds=10 → 200ms
Found: session 4 — OAuth added token refresh
Found: overnight — test showed p95 at 340ms
Root cause: bcrypt rounds=10 + OAuth token
refresh = 340ms on cold login.
Fix: reduce to rounds=8 (still secure),
cache OAuth tokens for 5min.
DAY 1
tenet init starts capturing your decisions and patterns.
WEEK 1
Suggestions reference previous work. Complex tasks build on what exists.
MONTH 1
Agents run overnight. Wake up to PRs that pass your evals.
MONTH 3
Ask about any past decision and get the full context with citations.

Your codebase improves while you sleep.

npm install -g @10et/cli

Free. Open source. Works with any AI model.

context that compounds