ainl-memory 0.1.1-alpha

AINL graph-memory substrate - agent memory as execution graph
Documentation

AINL Memory - Graph-Based Agent Memory Substrate

Graph-as-memory for AI agents. Execution IS the memory.

AINL Memory is a Rust library that implements agent memory as an execution graph. Every agent turn, tool call, and delegation becomes a typed graph node. No separate retrieval layer—the graph itself is the memory.

Why AINL Memory?

Most agent frameworks treat memory as separate from execution:

  • Execute → Store text → Retrieve → Execute again
  • Separate episodic, semantic, and procedural stores
  • Memory retrieval adds latency and complexity

AINL Memory unifies execution and memory:

  • Episodic memory: What happened during agent turns (tool calls, delegations)
  • Semantic memory: Facts learned with confidence scores
  • Procedural memory: Reusable compiled workflow patterns
  • Persona memory: Agent traits learned over time

All stored as typed graph nodes with edges for traversal.

Quick Start

[dependencies]
ainl-memory = "0.1.0-alpha"

use ainl_memory::{AinlMemoryNode, SqliteGraphStore, GraphStore};

// Open a graph store
let store = SqliteGraphStore::open("memory.db").unwrap();

// Record an agent delegation
let node = AinlMemoryNode::new_delegation_episode(
    "agent-A".to_string(),
    "agent-B".to_string(),
    "trace-123".to_string(),
    1, // depth
);

store.write_node(&node).unwrap();

// Query recent episodes
let episodes = store.query_recent_episodes("agent-A", 10).unwrap();

Features

  • Typed graph nodes: Episode, Semantic, Procedural, Persona
  • SQLite backend: Persistent graph storage with indexes
  • Graph traversal: Walk edges between nodes
  • Swappable backends: Implement GraphStore trait for custom storage
  • Zero unsafe code: Pure Rust implementation
  • Offline-first: No network required, local SQLite storage

Node Types

Episode

Records what happened during an agent turn:

  • Tool calls executed
  • Delegation to other agents
  • Trace ID for correlation
  • Delegation depth

Semantic

Facts learned with confidence:

  • Fact text
  • Confidence score (0.0-1.0)
  • Source turn ID

Procedural

Reusable workflow patterns:

  • Pattern name
  • Compiled graph (binary format)

Persona

Agent traits learned over time:

  • Trait name
  • Strength (0.0-1.0)
  • Source turn IDs

Architecture

AINL Memory is designed as infrastructure that any agent framework can adopt:

  • No dependencies on specific agent runtimes
  • Simple trait-based API
  • Bring your own storage backend

Status

Alpha (0.1.0-alpha)

API is subject to change. Production-ready for experimentation.

License

Licensed under either of:

at your option.

Contributing

Contributions are welcome! This is early-stage infrastructure—your feedback shapes the API.

See Also