aura-memory

Your AI model is smart. But it forgets everything after every conversation.

AuraSDK is a local cognitive runtime that runs alongside any frozen model. It gives agents durable memory, explainability, governed correction, bounded recall reranking, and bounded self-adaptation through experience — all locally, without fine-tuning or cloud training.

pip install aura-memory

from aura import Aura, Level

brain = Aura("./agent_memory")
brain.enable_full_cognitive_stack()  # activate all four bounded reranking overlays

# store what happens
brain.store("User always deploys to staging first", level=Level.Domain, tags=["workflow"])
brain.store("Staging deploy prevented 3 production incidents", level=Level.Domain, tags=["workflow"])

# recall — local retrieval with optional bounded cognitive reranking
context = brain.recall("deployment decision")  # <1ms, no API call

# inspect advisory hints produced from stored evidence
hints = brain.get_surfaced_policy_hints()
# → [{"action": "Prefer", "domain": "workflow", "description": "deploy to staging first"}]

No API keys. No embeddings required. No cloud. The model stays the same — the cognitive layer becomes more structured, more inspectable, and more useful over time.

⭐ If AuraSDK is useful to you, a GitHub star helps us get funding to continue development from Kyiv.

Why Aura?

The Core Idea: Cheap Model + Aura > Expensive Model Alone

Fine-tuning costs thousands of dollars and weeks of work. RAG requires embeddings and a vector database. Context windows are expensive per token.

Aura gives you a third path: a local cognitive runtime that accumulates structured experience between conversations — free, local, sub-millisecond.

Week 1: GPT-4o-mini + Aura                Week 1: GPT-4 alone
  → average answers                          → average answers

Week 4: GPT-4o-mini + Aura                Week 4: GPT-4 alone
  → recalls your workflow                    → still forgets everything
  → surfaces patterns you repeat             → same cost per token
  → exposes explainability + correction      → no improvement
  → boundedly adapts from experience         → no durable learning
  → $0 compute cost                          → still billing per call

The model stays the same. The cognitive layer gets stronger. That's Aura.

Performance

Benchmarked on 1,000 records (Windows 10 / Ryzen 7):

Mem0 recall requires an embedding API call (~200ms+) + vector search. Aura recall is pure local computation.

What Ships Today

Aura's full cognitive recall pipeline is active and bounded:

Record → Belief (±5%) → Concept (±4%) → Causal (±3%) → Policy (±2%)

Enable everything in one call:

brain.enable_full_cognitive_stack()   # activates all four bounded reranking phases
brain.disable_full_cognitive_stack()  # back to raw RRF baseline

Or configure individual phases:

brain.set_belief_rerank_mode("limited")   # belief-aware ranking
brain.set_concept_surface_mode("limited") # concept annotations + bounded concept reranking
brain.set_causal_rerank_mode("limited")   # causal chain boost
brain.set_policy_rerank_mode("limited")   # policy hint shaping

Higher layers also expose advisory surfaced output:

• get_surfaced_concepts() — stable concept abstractions over repeated beliefs

• get_surfaced_causal_patterns() — learned cause→effect patterns

• get_surfaced_policy_hints() — advisory recommendations (Prefer / Avoid / Warn)

• no automatic behavior influence — all output is advisory and read-only

Aura also ships operator-facing and plasticity-facing surfaces:

• explainability:

  • explain_recall()

  • explain_record()

  • provenance_chain()

  • explainability_bundle()

• governed correction:

  • targeted retract/deprecate APIs

  • persistent correction log

  • correction review queue

  • suggested corrections without auto-apply

• bounded autonomous plasticity:

  • capture_experience()

  • ingest_experience_batch()

  • maintenance-phase integration

  • anti-hallucination guards

  • plasticity risk scoring

  • purge / freeze controls

• bounded v6 cognitive guidance:

  • salience:

    • mark_record_salience()

    • get_high_salience_records()

    • get_salience_summary()

  • reflection:

    • get_reflection_summaries()

    • get_latest_reflection_digest()

    • get_reflection_digest()

  • contradiction and instability:

    • get_belief_instability_summary()

    • get_contradiction_clusters()

    • get_contradiction_review_queue()

  • honest explainability support:

    • unresolved-evidence markers in recall explanations

    • bounded answer-support phrasing for agent / UI layers

How Memory Works

Aura organizes memories into 4 levels across 2 tiers. Important memories persist, trivial ones decay naturally:

CORE TIER (slow decay — weeks to months)
  Identity  [0.99]  Who the user is. Preferences. Personality.
  Domain    [0.95]  Learned facts. Domain knowledge.

COGNITIVE TIER (fast decay — hours to days)
  Decisions [0.90]  Choices made. Action items.
  Working   [0.80]  Current tasks. Recent context.

SEMANTIC TYPES (modulate decay & promotion)
  fact          Default knowledge record.
  decision      More persistent than a standard fact. Promotes earlier.
  preference    Long-lived user or agent preference.
  contradiction Preserved longer for conflict analysis.
  trend         Time-sensitive pattern tracked over repeated activation.
  serendipity   Cross-domain discovery record.

One call runs the lifecycle — decay, promotion, consolidation, and archival:

report = brain.run_maintenance()  # background memory maintenance

Key Features

Core Cognitive Runtime

• Fast Local Recall - Multi-signal ranking with optional embedding support

• Two-Tier Memory — Cognitive (ephemeral) + Core (permanent) with decay, promotion, and archival

• Semantic Memory Types — 6 roles (fact, decision, trend, preference, contradiction, serendipity) that influence memory behavior and insighting

• Phase-Based Insights — Detects conflicts, trends, preference patterns, and cross-domain links

• Background Maintenance — Continuous memory hygiene: decay, reflect, insights, consolidation, archival

• Namespace Isolation — namespace="sandbox" keeps test data invisible to production recall

• Pluggable Embeddings - Optional embedding support: bring your own embedding function

Trust & Safety

• Trust & Provenance — Source authority scoring: user input outranks web scrapes, automatically

• Source Type Tracking — Every memory carries provenance: recorded, retrieved, inferred, generated

• Auto-Protect Guards — Detects phone numbers, emails, wallets, API keys automatically

• Encryption — ChaCha20-Poly1305 with Argon2id key derivation

Adaptive Memory

• Feedback Learning — brain.feedback(id, useful=True) boosts useful memories, weakens noise

• Semantic Versioning — brain.supersede(old_id, new_content) with full version chains

• Snapshots & Rollback — brain.snapshot("v1") / brain.rollback("v1") / brain.diff("v1","v2")

• Agent-to-Agent Sharing — export_context() / import_context() with trust metadata

Enterprise & Integrations

• Multimodal Stubs — store_image() / store_audio_transcript() with media provenance

• Prometheus Metrics — /metrics endpoint with 10+ business-level counters and histograms

• OpenTelemetry — telemetry feature flag with OTLP export and 17 instrumented spans

• MCP Server — Claude Desktop integration out of the box

• WASM-Ready — StorageBackend trait abstraction (FsBackend + MemoryBackend)

• Pure Rust Core — No Python dependencies, no external services

Advisory Cognitive Overlays

• Belief-Aware Recall Rerank — bounded production influence with strict guardrails

• Concept Overlay — surfaced concepts, per-record annotations, and optional bounded concept-aware reranking

• Causal / Policy Overlays — advisory surfaced output only, no automatic control path

• Cross-Namespace Analytics — read-only digest for tags, concepts, structural overlap, and canonical causal signatures across namespaces

Explainability & Governed Adaptation

• Explainability APIs — explain_recall(), explain_record(), provenance_chain(), explainability_bundle()

• Correction Governance — correction log, correction review queue, suggested corrections, namespace governance status

• Autonomous Cognitive Plasticity — extraction → ingest → maintenance loop for bounded self-adaptation without changing model weights

• Plasticity Safety Bounds — generated-confidence ceiling, risk throttling, purge/freeze controls, operator-visible risk state

Cognitive Guidance

• Salience Layer — bounded significance weighting for preservation, reminders, ranking, and operator review

• Maintenance Reflection — bounded reflection summaries and digests produced inside the normal maintenance pipeline

• Contradiction Governance — instability summaries, contradiction clusters, and bounded operator review queues for unresolved evidence

• Honest Answer Support — non-anthropomorphic phrasing hints for significance, uncertainty, contradiction, and reflection context

Quick Start

Trust & Provenance

from aura import Aura, TrustConfig

brain = Aura("./data")

tc = TrustConfig()
tc.source_trust = {"user": 1.0, "api": 0.8, "web_scrape": 0.5}
brain.set_trust_config(tc)

# User facts always rank higher than scraped data in recall
brain.store("User is vegan", channel="user")
brain.store("User might like steak restaurants", channel="web_scrape")

results = brain.recall_structured("food preferences", top_k=5)
# -> "User is vegan" scores higher, always

Pluggable Embeddings (Optional)

from aura import Aura

brain = Aura("./data")

# Plug in any embedding function: OpenAI, Ollama, sentence-transformers, etc.
from sentence_transformers import SentenceTransformer
model = SentenceTransformer("all-MiniLM-L6-v2")
brain.set_embedding_fn(lambda text: model.encode(text).tolist())

# Now "login problems" matches "Authentication failed" via semantic similarity
brain.store("Authentication failed for user admin")
results = brain.recall_structured("login problems", top_k=5)

Without embeddings, Aura continues to use its local recall pipeline - still fast, still effective.

Encryption

brain = Aura("./secret_data", password="my-secure-password")
brain.store("Top secret information")
assert brain.is_encrypted()  # ChaCha20-Poly1305 + Argon2id

Semantic Memory Types

brain = Aura("./data")

# Decisions are treated as higher-value memory
brain.store("Use PostgreSQL over MySQL", semantic_type="decision", tags=["db"])

# Preferences persist longer than generic working notes
brain.store("User prefers dark mode", semantic_type="preference", tags=["ui"])

# Contradictions are preserved for conflict analysis
brain.store("User said vegan but ordered steak", semantic_type="contradiction")

# Search by semantic type
decisions = brain.search(semantic_type="decision")

# Cross-domain insights surface higher-level patterns
insights = brain.insights(phase=2)
# Example:
# [{'insight_type': 'preference_pattern', 'description': 'Preference cluster around ui', ...}]

Namespace Isolation

brain = Aura("./data")

brain.store("Real preference: dark mode", namespace="default")
brain.store("Test: user likes light mode", namespace="sandbox")

# Recall only sees "default" namespace — sandbox is invisible
results = brain.recall_structured("user preference", top_k=5)

Cross-Namespace Digest

Use this when you need inspection-only analytics across isolated namespaces without changing recall behavior.

brain = Aura("./data")

digest = brain.cross_namespace_digest(
    namespaces=["default", "sandbox"],
    top_concepts_limit=3,
)

# Top concepts per namespace
print(digest["namespaces"][0]["top_concepts"])

# Pairwise overlap
print(digest["pairs"][0]["shared_tags"])
print(digest["pairs"][0]["shared_concept_signatures"])
print(digest["pairs"][0]["shared_causal_signatures"])

HTTP server:

GET /cross-namespace-digest?namespaces=default,sandbox&top_concepts_limit=3

MCP tool:

{
  "tool": "cross_namespace_digest",
  "arguments": {
    "namespaces": ["default", "sandbox"],
    "top_concepts_limit": 3
  }
}

The digest is read-only. It does not bypass namespace isolation in recall and does not feed training or inference by default.

For richer operator-facing workflows, see examples/V3_OPERATOR_WORKFLOWS.md.

Autonomous Cognitive Plasticity

Aura can also observe model output and feed bounded experience back into the cognitive substrate, without retraining the model.

from aura import Aura

brain = Aura("./data")
brain.set_plasticity_mode("limited")

capture = brain.capture_experience(
    prompt="How should we deploy this release?",
    retrieved_context=[],
    model_response="Deploy to staging first, then verify health checks before production.",
    session_id="deploy-session-1",
    source="model_inference",
)

brain.ingest_experience_batch([capture])
brain.run_maintenance()  # queued experience enters the normal cognitive pipeline

This stays bounded and operator-visible:

• the model remains frozen

• generated claims stay capped and guarded

• adaptation can be inspected, restricted, purged, or frozen per namespace

Recent operator HTTP endpoints:

• GET /explain-record

• GET /explain-recall

• GET /explainability-bundle

• GET /correction-log

• GET /cross-namespace-digest

• GET /memory-health

• GET /belief-instability

• GET /policy-lifecycle

• GET /correction-review-queue

• GET /suggested-corrections

• GET /namespace-governance-status

Cookbook: Personal Assistant That Remembers

The killer use case: an agent that remembers your preferences after a week offline, with zero API calls.

See examples/personal_assistant.py for the full runnable script.

from aura import Aura, Level

brain = Aura("./assistant_memory")

# Day 1: User tells the agent about themselves
brain.store("User is vegan", level=Level.Identity, tags=["diet"])
brain.store("User loves jazz music", level=Level.Identity, tags=["music"])
brain.store("User works 10am-6pm", level=Level.Identity, tags=["schedule"])
brain.store("Discuss quarterly report tomorrow", level=Level.Working, tags=["task"])

# Simulate a week passing — run maintenance cycles
for _ in range(7):
    brain.run_maintenance()  # decay + reflect + consolidate + archive

# Day 8: What does the agent remember?
context = brain.recall("user preferences and personality")
# -> Still remembers: vegan, jazz, schedule (Identity, strength ~0.93)
# -> "quarterly report" decayed heavily (Working, strength ~0.21)

Identity persists. Tasks fade. Important patterns get promoted. Like a real brain.

MCP Server — Claude Desktop · Cursor · Zed · VS Code

Give any MCP-compatible AI persistent, self-organizing memory:

pip install aura-memory

Claude Desktop — Settings → Developer → Edit Config:

{
  "mcpServers": {
    "aura": {
      "command": "python",
      "args": ["-m", "aura", "mcp", "C:\\Users\\YOUR_NAME\\aura_brain"]
    }
  }
}

Cursor / VS Code — .cursor/mcp.json or .vscode/mcp.json:

{
  "servers": {
    "aura": {
      "command": "python",
      "args": ["-m", "aura", "mcp", "./aura_brain"],
      "type": "stdio"
    }
  }
}

macOS / Linux path:

python -m aura mcp ~/aura_brain

Once connected, Claude automatically has 11 tools:

After connecting, tell Claude: "Before answering, always recall relevant context from memory. After our conversation, store key facts."

Windows test note

If cargo test intermittently fails on Windows with LNK1104 for target\debug\deps\aura-...exe, a stale test process is usually holding the file open. Run:

powershell -ExecutionPolicy Bypass -File .\scripts\cleanup_windows_test_lock.ps1

Then rerun the test command.

Dashboard UI

Aura includes a standalone web dashboard for visual memory management. Download from GitHub Releases.

./aura-dashboard ./my_brain --port 8000

Features: Analytics · Memory Explorer with filtering · Recall Console with live scoring · Batch ingest

Integrations & Examples

Try now: — zero install, runs in browser

FFI (C/Go/C#): aura.h · go/main.go · csharp/Program.cs

More examples: basic_usage.py · encryption.py · agent_memory.py · edge_device.py · maintenance_daemon.py · research_bot.py

Architecture

Aura uses a Rust core with Python bindings and a local-first memory runtime.

Publicly documented concepts are:

• Two-tier memory: cognitive + core

• Semantic roles for records

• Local multi-signal recall

• Belief-aware bounded reranking

• Trust, provenance, and namespace isolation

• Maintenance, insights, consolidation, and versioning

Higher cognitive layers may be present in the SDK as bounded reranking overlays and advisory inspection surfaces. They are not default runtime decision-making or behavior control.

The public repository documents the user-facing behavior and integration surface. Detailed internal architecture, tuning, and research notes are intentionally not published.

Resources

• Demo Video (30s) — Quick overview

• Examples — Ready-to-run scripts

• Landing Page — Project overview

Contributing

Contributions welcome! See CONTRIBUTING.md for setup instructions and guidelines, or check the open issues.

⭐ If Aura saves you time, a GitHub star helps others discover it and helps us continue development.

License & Intellectual Property

• Code License: MIT — see LICENSE.

• Patent Notice: Core architectural concepts are Patent Pending (US Provisional Application No. 63/969,703). See PATENT for details. The SDK source code is available under MIT. Separate commercial licensing is available for organizations that want contractual rights around patented architecture, OEM embedding, enterprise deployment, or dedicated support.

• Commercial Licensing: If you want to embed Aura's architecture into a commercial product, see COMMERCIAL.md.