Memex: Sovereign Memory for Autonomous Agents

Two layers. One vision. Memory you own.

---

The Vision

You will have agents. The question is: whose?

Option A: Rent bots from corps. Black boxes. Metered memory. Your identity, subscribed.

Option B: Build bots you control. Open architecture. Sovereign memory. Keys you hold.

This is the spec for Option B.

---

Two Layers

Layer 1: Memex (Memory Storage Engine)

The persistence layer. How agents remember.

• Observations — Time-indexed experiences with dual embeddings
• Contexts/Manifolds — Warped subspaces where different versions of you live
• Decay — Forgiveness-gated forgetting with anchor preservation
• Retrieval — Manifold-aware search that surfaces the right memories for right now

📁 memex/

Layer 2: Geometric Trust (Access Control)

The crypto layer. How agents share without surrendering.

• Transform Keys — Orthogonal rotations that warp query space
• Trust Graphs — Shared keys between agents enable mutual search
• Identity as Reachability — You prove who you are by reaching specific manifolds
• Permissions as Geometry — Wrong key = noise, not "access denied"

📁 geometric-trust/

---

The Core Insight

Memory isn't a warehouse. It's a space you inhabit.

Identity isn't "who you are." It's "which you is asking."

Permissions aren't gates. They're geometry.

Same key    → similarity preserved → FOUND
Different key → similarity destroyed → NOISE

The math does the access control. No ACLs. No tokens. Just: can you search that space or not?

---

Quick Start

# Layer 1: Store and recall memories
from memex import MemoryStore, StorageConfig

store = MemoryStore(StorageConfig(db_path="./agent.db"))

obs_id = store.observe(
    content="Found the brass key",
    agent_id="agent_1",
    timestamp=100
)

# Enter a context — queries get warped
store.enter_context("agent_1", "work")
memories = store.recall("key", agent_id="agent_1")

# Layer 2: Geometric access control
from geometric_trust import generate_key, transform_query

my_key = generate_key(dim=128)
query_warped = transform_query(query_vec, my_key)

# Only memories stored with my_key will correlate
# Everything else is noise

---

Files

memex-bundle/
├── README.md              # This file
├── ARCHITECTURE.md        # Full system design
├── memex/
│   ├── SPEC.md           # Full API specification (40+ pages)
│   ├── PLAN.md           # Implementation roadmap (9 phases)
│   └── pyproject.toml    # Package config
└── geometric-trust/
    ├── CONCEPT.md        # Permissions as geometry
    └── geometric_access.py  # Validation code (run it!)

---

Validated

True similarity:              0.9578
Correct key recovers:         0.9578
Wrong key sees:               0.0014 (noise)

Run python geometric-trust/geometric_access.py yourself.

---

The Challenge

$1,000,000 in prizes. Live streamed battles. Winner takes all.

Build your agent. Submit your application. Watch it fight.

Track	Prize Pool
Core Engine	$400,000
Manifold Math	$250,000
Edge Deployment	$200,000
Wild Card	$150,000

#MemexChallenge

---

License

CC BY-SA 4.0. Share it. Fork it. Warp it. Keep it open.

---

Memory as a service is a trap. Memory as geometry is freedom.