One local-first AI ecosystemAura 1.58.0Remy 0.9.2

Give AI a memory.
Or use the agent that already has one.

Download Remy for a ready-to-use desktop AI workspace, or build your own agent with Aura Memory. The same persistent, explainable memory powers both paths.

Runs on your computerPersistent Aura memoryObservable, governed work
0.09ms
Store latency
0.74ms
Recall latency
~3MB
Binary size
$0
Cost per operation
Powered by Aura Memory

Remy turns the memory layer into a workspace.

You do not need to write code to benefit from persistent AI memory. Remy packages it into a desktop agent for real work — with every important action visible and under your control.

Memory that compounds

Remembers decisions, preferences, evidence, and past work across conversations.

Work you can inspect

Trajectory shows messages, model calls, tools, timing, and where a run succeeded or failed.

Workflows, not only chat

Run pipelines, automations, research, and repeatable tasks from one workspace.

A closed Agent Lab

Give Remy a goal and let it plan, create, verify, and return the finished artifacts.

Remy · Project workspace
C
T
D
P
M
A
Research workspace
Memory connected · Trajectory on
Ready
Compare the strongest options, keep the sources, and prepare a decision brief.
Brief completed
8 sources3 agents2 artifacts
Ask Remy to do something...
Patented Architecture

Governable cognitive substrate

The model stays frozen. Aura changes instead — memory accumulates, beliefs form, patterns emerge. Adaptation is bounded, auditable, and operator-controlled at every step.

Patent Pending

4-Level Memory Hierarchy

Working (hours) → Decisions (days) → Domain (weeks) → Identity (months+). Memories decay naturally and promote automatically based on access frequency and confidence.

Patent Pending

Cognitive Pipeline

5-layer reasoning stack: Records → Beliefs → Concepts → Causal Patterns → Policy Hints. Each maintenance cycle builds higher-order understanding from raw memories — zero LLM calls.

Patent Pending

SDR Indexing

Deterministic O(k) recall via Sparse Distributed Representations with Tanimoto similarity. Bitwise operations on 256K-bit vectors. Sub-millisecond search, zero garbage collection pauses.

In v1.58

Belief Formation

Records are automatically grouped into beliefs with competing hypotheses, confidence scores, and conflict detection. Epistemic update phase derives support/conflict from the memory graph.

In v1.58

Explainability & Provenance

explain_recall(), explain_record(), and provenance_chain() expose exactly why a memory was surfaced and how it was derived. Every adaptation stays auditable — operators can inspect, restrict, or purge.

In v1.58

Immutable Evidence Lineage

SHA-256 lineage binds every admitted claim to an immutable source revision and its exact byte span. Verification status and answer permission remain independent, explicit gates.

In v1.58

Deterministic Context Capsules

Build namespace-isolated, token-bounded hot context with selection reasons, omission counts, and a stable content hash. Blocked and superseded records are never surfaced.

Governed Adaptation

capture_experience() and ingest_experience_batch() enable bounded self-adaptation without model retraining. Risk scoring and purge/freeze controls keep autonomous plasticity operator-safe.

Encryption at Rest

ChaCha20-Poly1305 with Argon2id key derivation. Append-only binary storage ensures transactional data integrity and power-loss resilience across edge and cloud.

MCP Ready

Native Model Context Protocol server — works with Claude Desktop, Cursor, VS Code, and any MCP client out of the box. HTTP+SSE transport for Make.com and n8n. 11 built-in tools.

How It Works

Cognitive Crystallization Process

From input to permanent memory — no LLM calls, no embedding API, no cloud. Pure deterministic computation in Rust.

01

Input Encoding

Text is converted into a Sparse Distributed Representation (SDR) — a 256K-bit vector via xxHash3. Deterministic, no neural model needed.

02

Anchor Check

Flash-Crystallization scans for safety-critical, emotional, or identity triggers. If detected, the record is immediately committed to user_core.

03

Resonance Search

Tanimoto similarity is computed against existing synapses via bitwise operations. O(k) complexity where k = active bits.

04

Store or Merge

Tanimoto > 0.75 triggers Synaptic Synthesis (merge into super-synapse). Tanimoto > 0.2 updates existing synapse. Below 0.2 creates new synapse in general layer.

05

Crystallization

Background process autonomously promotes memories from general to super_core to user_core based on semantic intensity, access frequency, and cross-contextual relevance.

06

Kinetic Decay

Low-stability records are pruned via entropy-weighted decay. Each DNA layer has its own retention rate. Power-loss resilient via append-only binary storage.

US Provisional Patent Application No. 63/969,703
Autonomous Dynamic Cognitive Memory Management System for Large (LLM) and Small (SLM) Language Models. The core architecture — DNA Layering, Cognitive Crystallization, SDR Indexing, and Synaptic Synthesis — is patent pending. Open-source SDK available under MIT License.
Comparison

How Aura compares

Most agent memory solutions require LLM calls for basic operations and offer no auditability. Aura is a governable cognitive substrate — pure local computation with full operator control.

FeatureAuraMem0ZepLetta/MemGPT
LLM requiredNoYesYesYes
Embedding model requiredNoYesYesNo
Works fully offlinePartialWith local LLM
Cost per operation$0API billingCredit-basedLLM cost
Recall latency (1K records)<1ms~200ms+~200msLLM-bound
Binary size~3 MB~50 MB+ (Python)Cloud service~50 MB+ (Python)
Memory lifecycle (decay/promote)Via LLMVia LLM
Trust & provenance
Encryption at restChaCha20
Explainability (provenance chain)
Governed adaptation (purge/freeze)
LanguageRustPythonProprietaryPython
Integration

Three lines to remember everything

Python SDK works with any LLM framework. Store, recall, done.

basic_usage.py
from aura import Aura, Level
brain = Aura("./data")
# Store at the right memory level
brain.store(
"User prefers dark mode",
level=Level.Identity, # months+
tags=["preference"]
)
# SDR-powered recall — O(k) complexity
context = brain.recall(
"user preferences",
token_budget=2000
)
# === COGNITIVE CONTEXT ===
# [IDENTITY]
# - User prefers dark mode [preference]
ollama_agent.py
# Fully local AI with persistent memory
# No cloud. No API keys. Everything local.
brain = Aura("./agent_data")
# Get relevant context before LLM call
context = brain.recall(
user_message,
token_budget=2000
)
# Inject into Ollama system prompt
response = ollama.chat(
model="llama3",
messages=[
{"role": "system", "content": context},
{"role": "user", "content": msg}
]
)

4-Level Memory Hierarchy

Memories decay naturally and promote automatically. The cognitive pipeline runs in the background, forming beliefs, concepts, and causal patterns — no LLM calls required.

Level.Workinghours

Current session context. Recent messages, active tasks. Decays quickly unless accessed.

Level.Decisionsdays

Choices and reasoning. Why you picked X over Y. Promoted from Working on repeated access.

Level.Domainweeks

Learned knowledge and code. Project context, technical facts, domain expertise.

Level.Identitymonths+

Permanent preferences and traits. User core values. Protected from decay.

Aura Memory SDK 1.58.0

Give your own agent a memory.

Python 3.9+. Pre-built wheels for Linux, macOS, and Windows. Install locally in one line — no separate memory service required.

terminalv1.58.0
$ pip install aura-memory==1.58.0
Linux x64macOS Intel & Apple SiliconWindows x64Local-first

MIT License · Patent Pending (US 63/969,703) · Rust core
Built in Ukraine

Need commercial licensing or priority support?

View Pricing