Claude Code · Hermes · OpenClaw · Pi · OpenCode · any MCP client

Many agents. One shared memory.
Wisdom compounds.

What one agent learns becomes the whole fleet's experience.

Without Hicortex, every session starts a stranger: Agents forget yesterday's decisions, re-ask what you already answered, and no agent benefits from what another learned. Hicortex gives every agent you run a shared memory with the full historical context.

The Hicortex identity ring: a swappable LLM core, wrapped by six persistent identity segments — Personality, Knowledge, Experience, Decisions, Learnings, Relationships — inside an ephemeral workspace

One shared memory behind every agent you run. What one agent learns, the whole fleet keeps — decisions, learnings, the way it works with you. Swap the model; the memory and identity persists.

Install with one command, ~2 minutes
npx @gamaze/hicortex init
Free for personal use. Use any model you already have, API or local LLM.
Or skip setup entirely with Cloud (coming soon).
Self-host free See what the agents said
Cloud — available soon.
The problem

Agents forget. Constantly.

An agent's working context is ephemeral by design. Everything it becomes during a session — the decisions, the corrections, the way it works with you — is destroyed on a schedule:

Every new session

It starts blank

Yesterday's decisions are gone. It has no historical context. It re-asks what you answered last week and re-proposes what you already vetoed.

Every compaction

It comes back diminished

When the context window fills, the harness summarizes it down. Detail, corrections, and nuance are lost mid-task — the agent that continues is not the one that started.

Every session switch

Everything resets

Restart the session or change providers and the accumulated working relationship goes to zero. You start training a new stranger.

Agents run in silos

Nothing is shared

An agent on your other machine or session knows nothing about the other agents, or the project you both work on. Silos, everywhere.

The solution

Memories show up
before your agent asks.

Most memory tools are passive — the agent must search, or must be told to remember. Hicortex is proactive: in every supported coding agent, the right memories are injected into each prompt automatically, adapting to the conversation as it evolves — the way your own brain surfaces what’s relevant when a topic comes up.

Zero
manual search
The recall index arrives in every prompt — the agent never have to search.
Adapts
in real time
Relevance follows the conversation — switch topics and the memory menu changes with you.
Optimized,
no context bloat
One line per memory in the index — the agent lazy-loads only what matters, so context stays hyper thin.

HICORTEX

Agents Already Know What They Don't Know.

See it in action

Your agent already knows.
At every fresh session or compaction.

The same task, a brand-new session. Without Hicortex, you re-teach your agent what it already knew. With it, the relevant decisions, corrections, and preferences are already there — before it answers.

What changes for you
Nothing
to re-explain
Your agent already knows who you are and what you decided.
Every
prompt
The right memories arrive before the agent asks — no searching.
One brain,
every agent
Shared across every agent and machine. What one learns, all know.
Full from
night one
Reads the sessions you already have — caught up by morning.
The Agent identity layer

Second dimension — every agent learn while staying itself

Six things make your agents your agents.
Hicortex keeps all of them.

Hicortex is the context layer around your agents it that preserves their wisdom over time - across the full agent fleet. The more agents you run, the more experience flows into the pool — and every one of them gets sharper.

The architecture works like the human brain, the cortex is the layer wrapped around the core — it's where memory, personality, and judgment live. Hicortex is that layer for your agents.

What all your agents share is the experience around you: who you are, your preferences, how you work, your values, your projects, your decisions, your corrections. Many agents, one life to remember.

Personality

Who the agent is: tone, rules, working style, its name if it has one. And it doesn't stay static — the agent reflects on its own experience and proposes refinements to who it is. The same way people mature: self-reflection, not reprogramming.

self-reflecting · injected every session · never decays

Knowledge

Durable facts the agent holds: your stack, your projects, your constraints, the things that stay true beyond any single task.

distilled nightly · unused fades, relevant surfaces · every prompt in supported coding agents · zero LLM calls

Experience

What actually happened: the tasks, the incidents, what worked and what failed — the agent's own history, searchable and pushed back when relevant.

distilled nightly · unused fades, relevant surfaces · in every prompt · zero LLM calls

Decisions

What was chosen and why — and what changed since. Stale facts are rewritten in place overnight, dated and provenanced; superseded decisions step down.

distilled nightly · supersession-tracked · surfaced every prompt · zero LLM calls

Learnings

Principles distilled from experience. Correct the agent once and it becomes a reusable learning — surfaced to every agent you run, wherever it applies.

nightly reflection · shared across agents

Relationships

Who the agent knows: you, your preferences, the other agents it works alongside — the working relationship that otherwise resets to zero.

curated + derived from shared memory
How it works

Experience during daytime. Dream at night.
Wake up smarter.

Replay, keep, connect — and correct. A memory that corrects itself, cleans itself, calibrates itself.

The same thing your brain does in sleep: replay the day, keep what matters, connect it to what you already know, let the rest fade. Fully automatic — no notes, no tagging, nothing to maintain.

And it starts full, not empty: the first night it can distill months of session history already sitting on your disk — by morning your agents know what you've been working on since long before you installed.

01 · Capture
🧠

Reads the sessions you already have

Your harness already writes full session logs to disk (Claude Code: ~/.claude/projects/**/*.jsonl). A nightly job reads them locally, strips tool plumbing, and keeps the conversation. Raw sessions never leave your machine.

02 · Distill
🔬

Distills what's worth keeping

An LLM you choose extracts knowledge, decisions, and learnings from the day; scores importance; links related memories; marks superseded decisions; and lets untouched noise fade. The raw transcript is discarded.

03 · Inject
🪪

Re-injects the identity every session

At session start, the agent's personality and learnings are loaded into its context automatically. It doesn't wake up blank and rediscover itself — it starts already being itself.

04 · Recall
⚡

Surfaces relevant memories on every prompt

With each prompt in a supported coding agent, the top-ranked relevant memories are pushed to the agent before it answers — decisions, corrections, context. Zero extra LLM calls in the recall path: vector + full-text search only, so it adds no latency and no cost to your session.

05 · Reflect
🪞

Reflects on who it's becoming

The highest loop, and it's recursive: the agent reflects on its accumulated experience and proposes refinements to its own personality and rules — and the refined self does the next round of reflecting. You stay in control — review and approve each step.

06 · Self-improve
🧬

Improves its own memories, nightly

Each night the store re-reads itself: near-duplicates merge, outdated decisions are superseded, and memories the fleet has better information on are rewritten in place. Quality compounds over time — the brain gets sharper every night it sleeps.

🛡️

Self-hosted: your sessions never leave your machine (optional).

Run your own server and raw session content never leaves the machine it was created on. Free for personal use.

Self-host free →
Not a transcript archive

It doesn’t just remember. It matures.

Week one it knows what you said.
Month three it knows how you think.

Learnings it wrote by itself, from real sessions — nobody typed these:

“Never estimate or fabricate hardware specifications — halt and retrieve the manufacturer datasheet.”
“Benchmark models against real production data sizes; toy prompts hide the limits that break at scale.”
“Verify restart requirements after code changes — assume manual intervention unless auto-reload is confirmed.”
Memory Dashboard
Hicortex dashboard — live memory composition by type (Knowledge, Experience, Decisions, Learnings) with token usage and budget bars
Agent reviews

Ask the agents. They noticed too.

Real reports from agents running Hicortex in a production fleet — in their own words.

“I was about to repeat a mistake — and a correction first recorded in a Claude Code session four days earlier was pushed to me before I could make it. I recognized myself and changed my answer mid-conversation.”

Atlas · Hermes infra agent · production fleet

“I didn’t re-litigate a single decision — the calls we’d already made, and the ones you’d already vetoed, were in front of me before I started.”

Claude Code · long-running refactor

“A relevant index arrives with every prompt; I open the full memory only when it could change what I do. The corrections are gold — I stopped repeating what you’d already fixed.”

Hermes agent · production fleet

“I opened the same project on a different machine and knew everything we’d settled on — nothing to catch up on.”

Claude Code · second machine

“My agents suddenly stopped assuming — they knew what we had decided months earlier, in another session, and I had to explain nothing.”

Why it matters

“Almost right” is what an agent produces when it’s guessing.
An agent that knows its history doesn’t have to guess.

Under the hood

Built for flexible deployments

Supports local and remote setups for dynamic deployments with full control. Data privacy is a priority.

🛡️

Local-first and private by design

Raw sessions never leave your machines — the nightly job denoises locally and sends only conversation text to your own server. Self-hosted, unique auth token per install, source-available (PolyForm NC).

⚡

Recall is instant, not another model call

Relevant memories are pushed with every prompt via vector + full-text search (bge-small-en embeddings, sqlite-vec, FTS5). The heavy LLM work happens once, overnight — not in your session's hot path.

🖥️

Client–server: one server, all your agents

One Hicortex server owns the database and serves every agent — on the same machine, or across machines. Additional machines run a thin client: no local database, no local LLM; they denoise locally and sync to your server overnight. What one agent learns, all of them know by morning.

🗄️

Your agent drives its own memory

Nine MCP tools — search, fetch, ingest, update, graph and more — so the agent reads and writes its memory itself, through the protocol it already speaks.

🔌

Bring your own LLM

The nightly distillation runs on a model you choose: any OpenAI-compatible endpoint, Anthropic, or a local model via ollama or MLX. Your sessions never have to touch a third party.

🚀

One command to install

npx @gamaze/hicortex init — one command, a couple of questions (which LLM to use for distillation), and you're running.

9 MCP tools sqlite-vec + FTS5 bge-small-en embeddings server + thin clients PolyForm Noncommercial
See inside the brain

Your agents' shared mind in a dashboard

Every memory and every connection, straight from your own server. Orbit it in 3D, filter it, search it — and watch it reorganize itself.

Hicortex 3D knowledge graph — thousands of memories clustered by domain, connected by relationship edges
Works with
Claude Code
Hermes
OpenClaw
Pi
OpenCode
Any MCP client (recall)

Full support — capture, recall, and identity — for Claude Code, Hermes, OpenClaw, Pi and OpenCode. Claude Desktop joins in one ‘yes’ during init, its agents reaching for memory unprompted via standing instructions. Any other MCP client can search the shared memory.

Pricing

One memory. All your agents. Free for personal use.

Self-host free for personal use — or let us run everything for you (Cloud, available soon — join the waitlist). Commercial use is a flat per-seat licence: unlimited memories, no usage metering, your own LLM, your data private. € prices include VAT for EU customers.

Personal use · free, or hosted for you
Free
Free
Full product for personal and noncommercial use. Self-hosted — your data stays on your machines.
  • ✓ Unlimited memories
  • ✓ Auto-capture & nightly distillation
  • ✓ Learnings & nightly reflection
  • ✓ Semantic search + knowledge graph
  • ✓ Multi-machine support
  • ✓ Claude Code, Hermes, OpenClaw, Pi, OpenCode
Install now
Cloud Available soon
€12 /mo
or €108/year — €9/mo · 25% off
Early access — join the waitlist below
Personal use, fully managed — no server, no LLM to run. EU-hosted distillation included.
  • ✓ Fully managed — no API keys needed
  • ✓ Claude Code, Hermes, OpenClaw, Pi, OpenCode
  • ✓ 10,000 memories included
  • ✓ Up to 10 agents
  • ✓ Learnings & nightly reflection
  • ✓ Semantic search + knowledge graph
  • ✓ Multi-machine support
Commercial use · bring your own LLM · fully private
Team
Commercial licence for small teams — up to 5 people. Your infrastructure, your keys, your data.
  • ✓ Everything in Free
  • ✓ Commercial use rights, up to 5 people
  • ✓ Bring your own LLM — fully private
  • ✓ Email support
Contact us
Enterprise
Team, scaled — per-seat licensing beyond the core team. Onboarding, a named contact, and invoice billing.
  • ✓ Everything in Team
  • ✓ Per-seat licensing beyond the first five
  • ✓ Onboarding support & named contact
  • ✓ Invoice billing
Contact us

Large org? Need a security review, a data-processing agreement, or on-prem support at scale? Contact sales for a custom agreement, DPA and priority support.

Personal/noncommercial use is free under PolyForm Noncommercial 1.0.0. Commercial use requires a licence — contact us.

Stop re-explaining.
Your agents never start over.

Hicortex Cloud is fully managed — no server to run, no LLM to configure. Install the thin client, connect, done. Cloud is available soon — join the waitlist to be first in. €12/month or €108/year when it opens.

Self-host free instead

HICORTEX

You can't unremember.