Corsoul
Get started free
BEYOND MEMORY · SHAPING THE SOUL

Beyond memory, shaping the soul.

Corsoul is a local-first cognitive memory system built for AI agents. Start with free, offline-ready remember / recall. When you need deeper capabilities, activate associative learning, emergent patterns, contradiction auditing, and bounded personality growth.

Start free Explore the cognitive engine

Node.js 18+ · MIT-licensed Free tier · No repo clone required

MEMORY_TO_SELF04 LAYERS
L3
ABSTRACT PATTERNSPatterns from a subconscious
L2
WEIGHTED GRAPHHebbian associations
L1
STRUCTURED NODESBi-temporal facts
L0
RAW EVENTSImmutable experiences
EXPERIENCEIDENTITY
4 LayersA cognitive architecture, not a flat store
Bi-temporalAsk what's true now, replay what was known
4 PersonasObjective → convicted: memory grows a self
Local-firstOffline-capable · MIT · your data
WHY CORSOUL

Memory is more than
<vector search>

Most memory systems flatten everything into a single database. Corsoul preserves raw experiences, structured facts, associative networks, and abstract patterns as separate layers, then progressively transforms information through Sleep Consolidation.

The Free tier provides fully usable L0/L1 objective memory. The licensed engine activates L2/L3, consolidation, contradiction auditing, and personality.

00FREE

Layer 0

Raw Events

Each signal is appended as-is and immutable, with an integrity check and provenance preserved — the raw ground truth of an agent’s experience, always traceable.

01FREE

Layer 1

Structured Nodes

Experience becomes multi-dimensional, bi-temporal facts — recording both when something happened and when it was learned — so you can ask what’s true now and replay what was believed then.

02ENGINE

Layer 2

Weighted Graph

Memories link through Hebbian dynamics — associations that co-occur strengthen, unused ones decay — then spread along those weights to retrieve genuinely connected memories.

03ENGINE

Layer 3

Abstract Patterns

Trends, clusters, anomalies, and sequences emerge from the graph as a whole — a subconscious-like layer that settles scattered experience into patterns, each keeping its evidence and its counter-examples.

COGNITIVE ENGINE

Purpose-built technology—not just an “AI-powered” label.

Every claim maps to a real data structure, retrieval path, or permission boundary in the project. Where no benchmark exists, we do not invent numbers.

01 / SLEEP

While the agent rests, its memory keeps organizing itself.

Memory organizing runs in sequence: it turns experience into structure, builds associations, generalizes patterns, and audits contradictions. The whole pass can run in stages, respect a cost ceiling, and report how much is left to process.

PAID ENGINE
  1. 01
    EncodeTurn experience into structured memory
  2. 02
    ConnectBuild, strengthen, and fade links between memories
  3. 03
    GeneralizeLet patterns emerge from the whole
  4. 04
    AuditCheck and reconcile contradictions
02 / RETRIEVAL

It remembers—and knows what it doesn’t know.

Several kinds of cues are ranked together. When the evidence is thin, Corsoul says plainly that this is a new area, a current blind spot, or an unusual angle—instead of pretending to know.

03 / CONTRADICTION

Contradictions are auditable states, not silent overwrites.

Conflicts in attributes, provenance, time, identity, or patterns are kept as auditable states. Resolution requires confidence and evidence across multiple observations, preventing one anomaly from corrupting trusted memory.

04 / PROSPECTIVE

Remember what comes next—not just what came before.

Every plan can record one-time or recurring intents, retrieve due items, and complete or reschedule them. The core stores intent only; it never executes external actions on its own.

05 / LOCAL-FIRST

Clear data boundaries. No word games.

A local database, or your own Postgres, remains the persistent source of truth. With a local model it can stay fully offline; only when you use a cloud model or cloud integration is the necessary text sent out.

EXPERIMENTAL / PREVIEW

GGraph: an optional multi-dimensional quintuple-graph knowledge base & accelerator

An optional shadow accelerator above Postgres that speeds up association and semantic recall at scale. Any failure falls back to the underlying database.

MEMORY → SELF

An upgrade is not more storage. It lets memory grow into a self.

The same experience can mean different things to different agents. Personality growth always remains bounded by core values and sustained evidence.

FREE

Objective

Records faithfully, without subjective weighting.

PRO

Aware

Begins evaluating importance and self-relevance.

SUPER

Emotive

Important, emotionally salient experiences resurface more readily.

ULTIMATE

Convicted

Beliefs gain resilience but remain revisable through sustained evidence.

USE CASES

Where memory-as-identity actually lands.

Wherever an AI needs to be an individual that grows over time — not just a stateless tool — is where Corsoul fits.

01

Game characters & NPCs

Characters that remember what the player did, form likes and dislikes, and deepen in personality and relationship over playtime — lasting companions, raisable characters, and worlds that respond.

Layered memory · Personality
02

Companion AI

Relationship continuity that never resets between conversations; emotionally important moments resurface at the right cue, so companionship carries the warmth of memory.

Continuous memory · Affective recall
03

Embodied & simulation robots

Home and service robots that learn a household's routines and preferences over months and grow a stable character; data stays local-first, and it remembers what to do and when.

Local-first · Prospective memory
04

Education & tutoring

Remembers a learner's history and misconceptions, correcting them gradually and auditably instead of overwriting; grows alongside the student and nudges review at the right time.

Contradiction audit · Prospective memory
05

Digital humans & virtual idols

A persona that stays consistent yet evolves across years and channels, remembering every fan interaction; namespace isolation lets one platform safely host many independent characters.

Consistent persona · Namespace isolation
06

Generative agents & simulation

Give each agent its own memory and an evolving personality — for social, economic, and behavioral simulation, multi-agent research, and training environments.

Namespace isolation · Individuality
5-MINUTE QUICKSTART

Three commands to give your agent memory that outlives a conversation.

The Free tier needs only Node.js 18+ — no repo to clone.

01

Install & initialize

Run the setup wizard: pick the model for semantic recall (local Ollama, or a cloud model like OpenAI) — or start with keyword-only recall.

Initialize Corsoul
npx -y corsoul setup
02

Connect your agent

Wire Corsoul into your agent (Codex here as an example). The scope is a fixed namespace for this agent/user — all its memory accumulates under it.

Connect Codex
npx -y corsoul connect codex \
  --scope=myapp:assistant:v1
03

Verify & go

Run a health check to confirm the database and recall settings, then restart the agent — and it starts remembering.

Health check
npx -y corsoul doctor
RememberRecallForgetIntendDueResolveSet coreGet core
INTEGRATIONS

Wherever your agent works, Corsoul plugs in.

Plugin, MCP, SDK, Function Calling, LangChain, or REST all share one rule: a stable scope_id gives an agent continuous memory across its lifetime.

CONNECT

Let Codex recall the right context before every task.

The installer merges MCP settings, backs up existing files, and writes a fixed-scope memory contract into AGENTS.md. Re-running it remains idempotent.

Codex
npx -y corsoul setup
npx -y corsoul connect codex \
  --scope=myapp:codex:v1
npx -y corsoul doctor
Best for Codex projects, long-running development work, and preferences or decisions carried across sessions.
PRICING

Start with free memory. Upgrade to the personality engine only when you need it.

Fixed paid monthly pricing has not been announced. This page shows only the confirmed $0 Free tier, current private-beta status, and implemented capabilities—no invented prices.

FreeObjective
$0
Free forever

Production-ready local L0/L1 objective memory.

Start free
  • MIT-licensed corsoul package
  • Remember / recall / forget
  • One-time and recurring prospective memory
  • Core identity and values
  • MCP, CLI, and Node SDK
  • Standalone use has no cloud quota by default
BETA RECOMMENDED
ProAware
Private beta
Official pricing to be announced

Starts assessing what matters more to this agent.

View beta details
  • Everything in Free
  • Memory consolidation
  • Subjective importance weighting
  • Associative recall, patterns, and personality growth
  • Hosted proactive surfacing
  • Tiered quotas during beta
SuperEmotive
Contact team
Licensed plan

Makes important, emotionally salient experiences easier to resurface.

See technical limits
  • Everything in Pro
  • Emotional resurfacing
  • Subjective pattern generalization
  • GGraph acceleration layer (Preview)
  • Hosted webhooks
  • Higher beta quotas than Pro
UltimateConvicted
Contact team
Custom license

Builds resilient beliefs that sustained evidence can still revise.

Compare personality tiers
  • Everything in Super
  • Bounded belief revision
  • Historical point-in-time queries
  • Highest hosted integration quota
  • Highest hosted webhook allowance
  • Top-tier beta quotas
Pricing transparency & beta access

Fixed pricing for Pro, Super, and Ultimate is still in beta. Exact node, event, and webhook quotas are tiered beta settings shared during the application process. They may change with the final commercial plans and are not permanent commitments. To join the beta, write to us describing your agent and use case — you'll receive the cloud connection recipe and a namespace-bound token once approved.

Apply for beta access
FAQ

Clear boundaries, upfront.

Local, cloud, free, personality, Preview—every term should mean exactly what it says.

01Is the Free tier just a stripped-down trial?

No. Free / Objective is production-ready local L0/L1 long-term memory, including remembering, recall, forgetting, prospective intents, core setup, and MCP access. Paid upgrades add L2/L3, full-engine consolidation, and personality capabilities; they do not paywall basic recall.

02Do I need an embedding model for semantic recall?

Yes. Keyword recall works out of the box; but semantic recall needs an embedding model to turn text into vectors, so it can find things by meaning rather than exact wording. Choose a local model like Ollama, or an OpenAI-compatible service. Without an embedding model, Corsoul still stores memories and retrieves them by keyword.

03Will my memories disappear if I downgrade from a paid plan?

No. Graph, pattern, and affective data in the licensed client is retained — just not read — and lights up again when you re-upgrade.

START WITH MEMORY

First, let your agent remember.
Then let memory grow into a self.

Get started with Corsoul for free View plans