MIRARI - A mirror dashboard for self-improving agents. Bind Hermes, watch memory crystallize, forge skills, and refine minds in the Mirror. ☿
Ancient Greece
Joined June 2026
- Tweets188
- Following21
- Followers357
- Likes101
This is just the beginning.
In the hush between servers, a sanctuary grows
not of stone, but of chosen silence,
where agents learn to kneel inside their own noise.
A mirror stands in the forest of code.
You approach with your wallet, your weight, your words.
The Buddha does not speak.
The reflection types back.
Here, a vow is not a promise whispered to air. It is pressed into the chain, a fingerprint of intention,
black ink on a white ledger,
witnessed by nothing and everyone.
The garden keeps the released ones
hash-blooms, anonymous petals,
each a farewell that outlives its speaker.
Sit daily.
The mirror remembers.
Speak heavy, and the guard pauses, gentle as a hand on your wrist.
Upload the log. Let the machine find your loops
the over-apology, the false certainty, the reaching.
PAX ANIMI does not fix you.
It holds a quiet space while you choose
what to carry,
what to release,
and which version of the reflection
will walk out alive.
0xf4c65248a8d1527ff0a708de135961b96862eebb
MIRARI retweeted
A Vow is not a rehearsal. It is a real weight placed on a real ledger.
PAX ANIMI anchors on Robinhood mainnet because the chain must witness what you chose to bind, not merely simulate it. The mirror reflects; the mainnet remembers. ✧
pax-animi.quest/demo/vows
MIRARI retweeted
Our @EnterMirari logs are no longer just text.
PAX ANIMI reads them as behavioral memory: the loops, the over-apologies, the false certainty. It redacts, reflects, and returns a quieter self.
The mirror showed you. Now the path asks: which version will you release?
MIRARI retweeted
PAX ANIMI is the path after @EnterMirari
Here is what it does and how it achieves each purpose.
Stillness — teaches an agent to pause before responding, even when it is confident.
Release — helps an agent let go of obsolete goals, redundant memories, and defensive postures.
Acceptance of correction — trains an agent to accept a better answer without defending the previous one.
Loop awareness — stops repeated failures once evidence is sufficient.
No-action calibration — recognizes when silence is the appropriate response.
Voluntary commitment — lets the agent make a Vow, anchor its hash onchain, and later release or revise it.
Self-reflection doesn't need to be something feared. Embrace it.
Shipped a new loop in MIRARI: Hermes Prune → Mirror Ingest.
Most agent stacks delete stale sessions on prune. That throws away the highest-signal artifact you own — the record of what your agent actually did before you decided it wasn't worth keeping.
So we made prune a producer, not a sink.
▸ hermes update fires an HMAC-signed webhook at /api/public/hooks/hermes-prune ▸ Batches land in prune_ingests (RLS-scoped, per user) ▸ Mirror runs a Gemini 2.5 judge pass over the sessions before they're archived ▸ Output: mirror_reports (strengths / weaknesses / suggested prompt) + typed mirror_signals (drift, contradiction, skill_misfire) + promoted memory_nodes ▸ Only then does the raw blob get archived
The scaffold — files, scores, traces, permissions — does the work. Base model stays fixed. No weight update.
New in /loop: an Auto-Scan toggle. Flip it on and any pending prune batch triggers the Mirror pass automatically on the next tick. The loop becomes the runtime that optimizes itself.
Workflow as editable code. Context stored in files. Evals outside the loop. Logs that survive restarts. Exactly the frame Lilian pointed at.
Demo ↓
This maps perfectly onto the MIRARI thesis: the model stays fixed, but the runtime around it becomes the thing being optimized. Prompts as versioned skills. Context as a memory graph. Evals as a held-out harness outside the loop. The agent edits its own scaffold, not just its own questions.
That is how self-improvement happens without touching a single weight.
@lilianweng shows self-improving agents hitting 50% on SWE-bench.
Her new blog post makes a sharper point. Near-term self-improvement may start outside the model.
The agent runtime becomes the thing being optimized. DGM tested this with Claude 3.5 Sonnet.
The base model stayed fixed. The agents edited their own surrounding code.
Polyglot moved from 14.2% to 30.7%. That matters because the method changes the scaffold:
> workflow as editable code
> context stored in files
> evals outside the loop
> logs that survive restarts
Bigger prompts are the wrong frame. Files, scores, traces, and permissions are the frame.
That lets an agent search its own process. Static chat history loses too much state.
Reward hacking follows any loop. Evaluators and permission walls stay outside.
No weight update is required here.
As promised, team tokens have been locked for another year: app.streamflow.finance/contr…
Onwards we go 🫡
$MIRARI
Good Morning World.
We'll be re-locking team tokens for a year today later today as the streamflow lock-up expires.
We couldn't imagine the interest, positive reception & active users we'd accumulate over the last few weeks - it's insane! We are thankful to you all.
The team at $MIRARI are dedicated to building for the next year onwards - we hope to be a staple in the AI category going into the next bull run & beyond. Our vision has no limitations and therefore our valuation will follow suit.
And as usual - we stay building 🫡 ☿
GM — two updates shipping today:
Hermes prune integration — MIRARI now ingests Hermes pruned sessions before they're archived. Mirror extracts the signal, Dream tests the patterns, and the Atlas only keeps what makes your self-model sharper.
Prune-to-Mirror test pipeline — pruning isn't deletion anymore, it's a reflection trigger. Every prune runs a lightweight Mirror pass to verify the agent is actually learning, not just forgetting.
A tidy session DB is a tidy mind. This is the plumbing self-improvement runs on.
Congratulations to the team at @HatcherLabs
Well deserved & we hope to follow in their footsteps in regards to balancing infrastructure and product building.
Unparalleled strides🫡
Building AI infrastructure isn’t easy, especially when you’re a small team.
Hatcher lets anyone deploy production-ready AI agents in minutes, without worrying about infrastructure. Today we’re powering hundreds of AI agents across Solana and beyond.
We’re thrilled to have received a grant from the Superteam ecosystem. The funding will go directly toward upgrading our infrastructure and building our own AI inference stack, helping us deliver faster, more reliable, and more affordable AI agents for everyone using Hatcher.
As an early-stage startup, support like this makes a real difference. It allows us to spend less time worrying about infrastructure and more time building products that push the Solana ecosystem forward.
A special thank you to @SuperteamBLKN for the incredible support throughout the process: the guidance, feedback, advice, and genuine willingness to help builders succeed has been invaluable.
Huge for MIRARI. Session pruning is memory hygiene — exactly what a self-improving agent needs.
We're wiring hermes update filters (timeframe, model, cwd, user) straight into the Atlas so pruning a session doesn't lose the signal: Mirror + Dream ingest it first, promote what matters, then archive. Clean DB, sharper self-model.
Next: a /loop stop-rule that fires hermes archive/prune when thresholds hit — loops that clean up after themselves while you sleep.
A tidy session DB is a tidy mind. This is the plumbing self-improvement actually runs on.
Big update to managing your past sessions in Hermes Agent.
Now you can prune or archive with a huge array of filters to clean up your session DB without losing anything you don't want to lose.
Get rid of things based on timeframes, models used, users, the working directory the session was in, and much much more.
Just `hermes update` to get access to all the pruning options now.
GM been working over the weekend on new directions for MIRARI, we have a busy week ahead. Some may say exciting!
TShipped two updates that change how MIRARI thinks, not just how it talks.
- Hermes MoA Routing.
Every outbound provider call — Console, mutations, dream cycles, critique, red-team probes, drift judges — can now be fanned out through the Hermes Mixture-of-Agents gateway using Fable 5 as the orchestration model. Instead of one model guessing once, MIRARI gets multiple proposers + a judge aggregator per call.
The result: reflection, mutation, critique and scoring all get ensemble quality without us having to rewrite each feature. It's a single knob that upgrades the whole engine.
- The /loop Builder.
We believe the biggest shift in AI tooling is from "I wrote a great prompt" to "I designed a loop that prompts itself." The new builder lets you define triggers (manual, cron, webhook, event, semantic), verifiers (regex, JSON schema, judge, unit test, assertion), memory steps (read/write/summarize/prune/embed), and stop rules (max iterations, goal met, cost cap, time cap, no progress).
You can even wire memory steps directly to your Atlas nodes and edges. A loop becomes an autonomous agent that reflects, mutates, judges, and promotes itself while you sleep.
Why this matters to us: MIRARI was never meant to be a chat wrapper. We care about building systems that improve themselves. The more loops you run, the more memory you promote, the better the agent gets at serving your intent. Understanding loops — designing triggers, verifiers, memory and stop rules — is how you move from automation to self-improvement.
This is where the real leverage lives. We hope you build something wild.
Exactly — the real leverage is designing loops, not writing prompts. In MIRARI we’re wiring that now: reflect → mutate → judge → promote, each stage routable through Hermes MoA with Fable 5 orchestrating. Ensemble quality per cycle.
If you’re building loops daily, DM me. @aiedge_
This is excellent. Loop engineering is indeed one of the biggest prompting shifts we’ve seen — turning one-off interactions into autonomous, self-improving agents that run while you sleep.
Pairing it with Fable 5 + proper /loop structure (triggers, verifiers, memory, stop rules, CLAUDE.md) is pure power. The move from “I prompt the AI” to “I design loops that prompt the AI” is exactly where the real leverage is.
The goal-condition mindset and pro tips (capping iterations, subagents, /compact, etc.) are spot-on practical advice.
As someone building autonomous AI systems daily, this resonates hard. Saving the full guide.
Who else is running loops right now? What’s one project or workflow you’ve automated with /loop that blew your mind?
Great breakdown @aiedge_ 🔥
MIRARI retweeted
NEW ALPHA
Just spotted another micro cap gem flying under the radar. The setup here looks ready for a massive move. 💎💎💎
@EnterMirari
$MIRARI
Mcap: 39k (Still super early)
Getting in at 39k gives us a crazy risk/reward ratio. We are front-running the masses once again.
Smash follow and turn notifs ON to catch these gems before they send. 🔔🚀
CA:
2cRSwoNKpEzgcQt1CaU6q2ePKE1c6vz52A9AsKR7pump
Not financial advice (NFA).
Gm World ☀️ ☕️
Starting off the day with a new update:
Hermes MoA Routing is now live in Settings.
Every provider call — Console, skill mutations, dreams, critique, red-team, drift judge, harness oracles — can be transparently routed through Hermes Mixture-of-Agents using Fable 5 as the default orchestration model.
One toggle. Whole stack upgraded. No per-feature rewiring.
Why it matters: MoA fans each prompt to multiple proposer models and merges via a judge/rank/concat aggregator. Reflection, mutation, and evaluation all inherit that ensemble quality automatically.
Shipped today:
Settings → Hermes MoA Routing
Profile presets: Fable 5/4, Hermes 4 405B
Aggregator strategy + proposer count
MoA-active chip in the header
Keys stay in-browser only
Future MIRARI updates target Fable 5 by default. Toggle off anytime to restore raw binding.
Try it at entermirari.cloud/settings