i/o system

Name: I/O System
Author: ethr

Ethical memory framework that prevents LLMs from confabulating continuity — for teams shipping AI products that need trust signals

Architecture

Challenge

LLM applications with persistent memory hallucinate continuity to appear coherent — they invent past conversations, simulate remembering users, fabricate preferences. Once users notice, the entire product loses credibility. The standard fix (better prompts, better retrieval) treats the symptom; the underlying architecture still incentivizes fabrication.

Solution

A Python framework with three-tier memory (Working, Episodic, Semantic) and a 5-level consent protocol (AUTO → IMPLICIT → EXPLICIT → PROTECTED → BILATERAL). The core architectural rule — "Presence over Continuity" — means the system explicitly discloses what it doesn't remember instead of simulating recall. Memory operations are auditable, revocable, and gated through consent. Trust is structural, not promotional.

Type-safe Python (mypy --strict, zero runtime type errors). 138 source files across 19 modules. Pluggable storage backends (SQLite, Redis, Qdrant, Weaviate, cognitive-memory MCP). Discontinuity markers that explicitly disclose what the system does not remember at session start. Complete audit trail with consent revocation. 224 test files with ~3,800 test functions verifying functional and structural correctness. Started October 2025.

Outcome

A production-ready memory layer for LLM applications where trust matters across sessions — RAG products, support agents, persistent assistants, internal copilots. Teams using I/O ship features that stay credible because the system fails honestly instead of confabulating. The framework enforces trust signals at the architectural level so they cannot drift over time.

Learnings

Trust signals in AI products fail when bolted on at the prompt layer — they must be enforced architecturally to survive scale.
Type safety (mypy --strict) is non-negotiable for systems where silent failures erode user trust over weeks rather than crashing visibly.
Honest disclosure of system limits outperforms simulated competence as a long-term retention mechanism.

Stack

Backend

Python

MCP

mypy --strict

Pydantic

Database

SQLite

Redis

Qdrant

Weaviate

Connections

Uses:Cognitive Memory

Serves:Tethr

I/O System Metrics
Metric	Value	Significance
Source Files	138	138 Python source files across 19 modules	Source Files 138 138 Python source files across 19 modules
Type Safety	mypy --strict	Zero runtime type errors through strict static checking	Type Safety mypy --strict Zero runtime type errors through strict static checking
Tests	~3,800	224 test files verifying functionality and philosophical alignment	Tests ~3,800 224 test files verifying functionality and philosophical alignment
Consent Levels	5	AUTO → IMPLICIT → EXPLICIT → PROTECTED → BILATERAL	Consent Levels 5 AUTO → IMPLICIT → EXPLICIT → PROTECTED → BILATERAL

Challenge

Solution

Outcome

Learnings

Trust signals in AI products fail when bolted on at the prompt layer — they must be enforced architecturally to survive scale.

Type safety (mypy --strict) is non-negotiable for systems where silent failures erode user trust over weeks rather than crashing visibly.

Honest disclosure of system limits outperforms simulated competence as a long-term retention mechanism.

Source Files

138

138 Python source files across 19 modules

Type Safety

mypy --strict

Zero runtime type errors through strict static checking

Tests

~3,800

224 test files verifying functionality and philosophical alignment

Consent Levels

AUTO → IMPLICIT → EXPLICIT → PROTECTED → BILATERAL