Aleutian Logo AleutianFOSS

Give Your AI Agent a Memory

Persistent graph indexes + small decision model = fantastic results on CPU

Stop rebuilding context every session. AleutianFOSS gives AI coding agents persistent understanding of your codebase using lightweight graph infrastructure and a small CPU-runnable model for decision-making. Six synchronized indexes. Zero GPUs required.

Explore on GitHub Quick Start Market Positioning

❌ Without AleutianFOSS

  • Context rebuilt every session
  • Token limits force amnesia
  • Large models for simple queries
  • Wasted compute on re-parsing
  • No persistent code understanding

✓ With AleutianFOSS

  • Persistent graph memory
  • Small model navigates structure
  • CPU-runnable decision making
  • 5000x cache speedups
  • Fantastic results, zero GPUs
CPU
No GPU Required
5000x
Graph Cache Speedup
6
Persistent Indexes
22x
Memory Reduction

The Problem: AI Agents Have Amnesia

Current AI coding agents rebuild context from scratch every session. This wastes compute, hits token limits, and forces you to use large expensive models for simple structural queries.

Example: Finding All Callers of a Function

❌ Typical AI Agent

  • 1. Read entire codebase into context
  • 2. Parse all files with large LLM
  • 3. Search for function calls
  • 4. Hit token limit, truncate
  • 5. Next session: repeat from step 1

Cost: $$$, slow, incomplete

✓ AleutianFOSS

  • 1. Query persisted Call Graph
  • 2. O(1) hash lookup in Symbol Table
  • 3. Small model formats results
  • 4. Sub-millisecond response
  • 5. Next session: same speed

Cost: ¢, instant, complete

The Insight: You don't need a large model to navigate code structure. You need persistent graphs and a small decision-making model. AleutianFOSS provides the graphs. You bring a CPU-runnable model (7B, 4B, even 1B works) to make decisions.

Lightweight Architecture

Six persistent graph indexes handle structure. Small CPU-runnable model handles decisions. No GPUs. No expensive inference. Fantastic results.

AleutianFOSS Architecture
Six Persistent Graph Indexes · Disk-Backed
Call Graph
Dominators
Import Graph
Modules
Symbol Table
O(1) lookup
Type Hierarchy
Inheritance
HLD Queries
O(log²V) LCA
Communities
Leiden
Structured Queries
Small Decision Model · CPU-Runnable
  • 7B, 4B, or even 1B parameters
  • Runs on CPU (no GPU needed)
  • Uses graph structure for context
  • Makes decisions, not structural queries
Agent Decision

Key insight: separate structure (graphs) from reasoning (model).
Graphs persist across sessions. Model stays small and fast.

Persistent Graphs

Six synchronized indexes stored in BadgerDB. Content-addressable caching (SHA256 hash). Build once per commit, query forever. Survives restarts, sessions, reboots.

Small Decision Model

Use 7B, 4B, or 1B parameter models on CPU. Graphs provide structure, model provides reasoning. No GPU inference costs. Fantastic results from small models + good data.

5000x Cache Speedup

Graph cache hit: 5s → 1ms. HLD session startup: 6s → 163ms. Parallel BFS memory: 11.2MB → 493KB. Production-optimized for real codebases.

Advanced Algorithms

Dominators (bottlenecks), PageRank (criticality), Leiden communities (modules), Heavy-Light Decomposition (O(log²V) path queries). All pre-computed and cached.

27 Built-In Tools

find_dominators, find_hot_spots, get_callers, find_lca, path_aggregate, and 22 more. Ready-to-use for your agent. OpenTelemetry instrumented.

AGPL v3 + AI Attribution

Open source with copyleft protection. AI systems must attribute usage. Prevents corporate LLMs from laundering your infrastructure work.

Lightweight = Fast

20+ algorithmic optimizations. Production-validated on 100K+ node graphs. All metrics from real codebases, not synthetic benchmarks.

Operation Without AleutianFOSS With AleutianFOSS Improvement
Graph cache hit 5000ms (rebuild from AST) 1ms (content-addressed) 5000x faster
Find symbol definition O(V) linear scan O(1) hash lookup 100x faster
Parallel BFS memory 11.2 MB (concurrent maps) 493 KB (optimized) 22x reduction
HLD session startup 6024ms (rebuild tree) 163ms (BadgerDB load) 37x faster
Community detection Every query recomputes Cache-first lookup 100x faster
LCA path queries O(V) tree traversal O(log² V) HLD 1000x on deep trees
Critical path analysis Multiple BFS passes Single traversal 3.5x faster
Dominator computation Full graph every time Subgraph + memoization 10-250x faster

Why This Matters: A small 4B model on CPU can outperform a 70B model on GPU for code navigation tasks—when it has the right structural data. AleutianFOSS provides that data. You save on inference costs, hardware, and get faster results.

Quick Start

Get AleutianFOSS running on your codebase in under 5 minutes. Currently supports Go with Python, JavaScript, and Rust in development.

Installation
# Clone the repository
git clone https://github.com/AleutianAI/AleutianFOSS.git
cd AleutianFOSS

# Build the trace agent (Go 1.21+)
cd services/trace/cli
go build -o trace .

# Initialize graphs on your codebase
./trace init /path/to/your/go/project

# Run queries
./trace query find_hot_spots --limit 10
./trace query find_dominators --threshold 5
./trace query get_callers --symbol "MyFunction"

Integration with Your Agent

Your small CPU-runnable model queries the graphs, AleutianFOSS returns structured data.

agent_integration.go
package main

import (
    "context"
    "github.com/AleutianAI/AleutianFOSS/services/trace/cli/tools"
    "github.com/AleutianAI/AleutianFOSS/services/trace/graph"
)

func main() {
    ctx := context.Background()

    // Initialize graphs (persisted across sessions)
    g := graph.NewCallGraph()
    analytics := graph.NewAnalytics(g)

    // Register tools
    dispatcher := tools.NewDispatcher()
    tools.RegisterExploreTools(dispatcher, analytics, g, nil)

    // Your small model decides what to query
    // AleutianFOSS provides structured results
    result, err := dispatcher.Execute(ctx, "find_hot_spots", map[string]interface{}{
        "limit": 10,
    })

    // result: Top 10 critical functions by PageRank
    // No LLM needed for structural queries
}

Pro Tip: Use a 7B quantized model (GGUF format, CPU-runnable) as your decision maker. It queries AleutianFOSS graphs for structure, focuses its reasoning on high-level decisions. Costs pennies, runs on a laptop, beats 70B models on code tasks.

Strategic Roadmap

Three-phase architecture: Code Understanding (graphs) → Planning (small model decisions) → Code Generation. Phase 1 is production-ready. Phase 2 in active development.

Phase 1: CRS ✓
Persistent Graphs

6 indexes, 27 tools, production-ready

Phase 2: Planning
Small Model Decisions

MCTS, beam search, active development

Phase 3: Code Gen
Implementation

Builds on graphs + planning

Research & Documentation

40+ technical articles documenting architecture, optimizations, and algorithm implementations.

Heavy-Light Decomposition
O(log² V) path queries on 100K+ node trees
Performance Optimizations
20 algorithmic improvements documented
Market Positioning
Strategic analysis and roadmap
All Articles
40+ technical deep dives

Ready to Build?

Join researchers and infrastructure teams building AI agents with persistent memory and CPU-runnable decision models. No GPUs required.

Star on GitHub Join Discord Back to Main Site