Skip to content

🧠 MemEX v0.9.0 — Initial Release

Choose a tag to compare

View all tags
@LZL0 LZL0 released this 12 Apr 05:28
· 86 commits to master since this release
0.9.0
ffca8ea

Multi-session continuity for AI systems.

MemEX is a structured memory substrate that represents knowledge as an evolving belief state, with explicit provenance, scoring, and contradiction-aware updates.

This is the first public release.

What MemEX is

Most AI systems treat memory as retrieved text.

MemEX treats memory as structured belief.

Instead of appending and summarizing chat logs, MemEX stores:

  • Observations, assertions, hypotheses, and policies
  • Provenance chains (what something was derived from)
  • Explicit contradictions (not silent overwrites)
  • Scores for authority, conviction, and importance

The result is a system that can:

  • carry forward state across sessions
  • reason about what it knows
  • explain why it believes something

What’s in 0.9.0

🧠 Belief-state memory graph

  • Typed MemoryItem model (observation, hypothesis, derivation, etc.)
  • Provenance via parents and support trees
  • First-class contradiction handling (CONTRADICTS, SUPERSEDES)
  • Multi-dimensional scoring (authority, conviction, importance)
  • Time decay applied at query time (no mutation of stored state)

🔍 Retrieval as a pipeline

  • Filter → score → decay → contradiction handling → diversity → budget
  • Scored retrieval with configurable weights
  • Smart retrieval with contradiction surfacing and diversity penalties
  • Budget-aware context packing

🎯 Intent & Task graphs

MemEX separates cognition into three layers:

  • Memory — what is believed
  • Intent — what is wanted
  • Task — what is done

Each layer has:

  • its own types
  • reducer + lifecycle events
  • explicit linking via IDs

Together they form a continuous loop:

Memory → Intent → Task → Memory

🔄 Transplant (export / import)

MemEX can export and import self-contained slices of cognitive state:

Memory (belief)
+ Intent (goals)
+ Task (execution history)
= Transferable cognitive state

This enables:

  • sub-agent isolation
  • parallel reasoning
  • safe delegation
  • cross-system state transfer

Import is append-only by default, with optional:

  • conflict detection
  • re-ID with reference rewriting

⏱️ Timestamp-preserving re-ID

When importing conflicting entities:

  • new IDs are generated at original_timestamp + 1ms (from UUIDv7)
  • not Date.now()

This preserves:

  • recency ordering
  • decay scoring
  • temporal semantics

📦 Event-driven architecture

  • Command → reducer → lifecycle events
  • Immutable state transitions
  • Replayable command logs
  • Envelope-based integration with external systems

Cognitive Transfer

MemEX supports transferring cognition between agents:

Agent A → Agent B:

  Memory  (what is known)
+ Intent  (what is being pursued)
+ Task    (what has been attempted)
= Continued execution without reset

The receiving agent:

  • inherits context
  • continues active goals
  • avoids previously failed approaches

It does not restart. It continues.

Design Principles

  • Structured over textual — memory is typed, not appended text
  • Explicit over implicit — contradictions, provenance, and confidence are first-class
  • Append-only — no silent overwrites
  • Separation of concerns — belief, goals, and execution are distinct
  • Minimal core — no storage, UI, or transport included

What this is not

MemEX is not:

  • a vector database
  • a chat memory wrapper
  • an agent framework

It is:

a memory substrate that other systems can build on

Status

0.9.0 is the first public release.

  • Core model is stable
  • APIs may still evolve
  • Intended for experimentation and system integration

Install

npm install @ai2070/memex

Closing

MemEX is an attempt to treat memory as a first-class system, not a side effect of prompting.

If you build on it, feedback is welcome.

Assets 2
Loading