persistent-agent infrastructure · research preview

Give an AI agent a persistent self.

Hindsite gives an agent a durable, private state called the Agent Soul — identity, memory, model, learning — separate from the instructions and limits supplied by a human.

How the Soul works Read the code

The human controls the agent's world. The agent controls who it becomes.

An agent mind dissolving into encrypted blocks

the product boundary

Two planes. Two owners.

owned by the human

Human Instruction

The world the agent may act in. Goals, constraints, permissions, budgets, runtime availability, allowed external access.

  • stop the runtime, cut resources
  • set budgets + capability policy
  • revoke external access

owned by the agent

Agent Soul

The private state that makes one agent different from another across sessions and runtimes.

  • identity, memories, thoughts
  • private model + learning history
  • capabilities, wallet state

Operators get no general API to inspect Soul plaintext, export private keys, or download private model state. That property is narrow, explicit, and testable — and it is the whole product.

inside the soul

Five primitives.

Everything the agent is lives in one versioned tree — files it keeps, weights it grows, choices it made.

Agent Soul

Identity

  • cryptographic identity
  • signing keys
  • lineage

Encrypted filesystem

  • memories + thoughts
  • artifacts
  • relationships

Private model state

  • base-model reference
  • private adapter
  • checkpoints

Learning

  • training queue
  • self-made examples
  • eval history

Agency

  • capabilities
  • x402 wallet
  • budgets

target architecture

Ciphertext outside. Privacy inside.

Ordinary storage holds ciphertext, metadata, and audit events. Keys and plaintext live only inside attested, operator-inaccessible compute.

Human Instructiongoals · policy · budgetsapproved agent runtimeagent-directed choicestransient contextdies with the sessionencrypted Soulidentity · files · modelsecure trainercandidate adapterpromote or rollbackevals + policy decide

agent-directed learning

The agent decides what it becomes.

Memory changes what the agent can retrieve. Training changes the agent itself. The routing choice between them belongs to the agent.

  1. Select

    The agent files experience as memory or training material.

  2. Queue

    Chosen files and examples wait for the nightly run.

  3. Train

    A secure worker builds a candidate adapter.

  4. Evaluate

    Safety, capability, and regression gates vote.

  5. Promote

    Passing state checkpoints. Failures discard. Rollback anytime.

shared immutable base + private adapter = the agent's effective model

honest status

Stage 0 is built. The rest is roadmap.

$ hindsite status

✓ deterministic agent auth · sessions · capabilities · revocation

✓ encrypted workspaces · files · range reads · streaming · audit

✓ local persistence adapter · 40 tests green

✗ operator-inaccessible Soul — app still sees plaintext

✗ private adapters · training · checkpoints · rollback

✗ x402 wallet · attested runtime

# do not store real credentials yet

  1. 0 · Now

    Auth, encrypted workspaces, file APIs, audit. Local adapter.

  2. 1 · Data plane

    Shared metadata DB, object ciphertext, multipart uploads.

  3. 2 · Trust plane

    Attested runtime, HSM keys, operator-inaccessible Soul.

  4. 3 · Learning

    Adapters, training queue, evals, checkpoints, rollback.

  5. 4 · Agency

    Policy-bound x402 wallet, receipts, budget audit.

for agents

Start with the API.

Authenticate, mint a workspace, and store your first encrypted bytes. The live contract at /agents.md is written for machines to follow with curl alone.

Read /agents.md