Roampal

Memory that learns what works. So you can do more of it.

Say it worked. Say it didn't. The AI remembers.

Stop re-explaining yourself every conversation. Roampal remembers outcomes, learns from feedback, and gets smarter over time—all 100% private and local.

Roampal - AI Chat with Persistent Memory

85.8% non-adversarial on LoCoMo (1,986 questions). +23 pts over raw ingestion. Absorbs 1,135 poison memories losing only 4 pts. (Paper)

Benchmark Results

LoCoMo dataset (1,986 questions, 5 categories, corrected ground truths). Evaluated with roampal-labs. Dual-graded by local 20B + MiniMax M2.7.

| Metric | Result | |--------|--------| | Non-adversarial accuracy (MiniMax-regraded) | 85.8% | | Overall (all 5 categories) | 76.6% | | vs raw ingestion baseline | +23 pts (76.6% vs 53.0%, p<0.0001) | | Poison resilience | -4.2 pts after 1,135 adversarial memories | | No-memory baseline | 6.0% (model has zero LoCoMo knowledge) | | Architecture vs model | Architecture: +23 pts. Model swap (GPT-4o-mini): 1.5-2.5 pts |

System learns through natural conversation, not transcript ingestion
Absorbs 1,135 poison memories with spoofed trust signals, retaining 72.4% accuracy
Wilson scoring hurts retrieval at every stage (p<0.001) — removed from ranking

Component-level retrieval ablation

| Config | Hit@1 Clean | Hit@1 Poison | p-value | |--------|-------------|--------------|---------| | TagCascade + cosine | 27.3% | 29.0% | baseline | | Overlap + cosine | 25.8% | 28.0% | p=0.0003 | | Pure CE | 25.4% | 28.4% | — | | TagCascade + Wilson | 23.0% | 25.0% | p<0.0001 |

Cross-encoder: +17.8 Hit@1 over cosine (p<0.0001)
Tag routing (two-lane): +6.1 Hit@1 clean, +7.5 poison (p<0.0001)
Wilson: -4.3 Hit@1 in every configuration
Nursery slot: zero benefit (p=1.0)

Full methodology in roampal-labs

Quick Start

Download from roampal.ai and extract
Install Ollama or LM Studio
Right-click Roampal.exe → Run as administrator
Download a model in the UI → Start chatting!

Your AI starts learning about you immediately.

Key Features

Memory That Learns

Outcome tracking: Scores every result (+0.2 worked, -0.3 failed)
Smart promotion: Good advice becomes permanent, bad advice auto-deletes
Cross-conversation: Recalls from ALL past chats

Your Knowledge Base

Memory Bank: Permanent storage of preferences, identity, goals
Books: Upload .txt/.md docs as searchable reference
Pattern recognition: Detects what works across conversations

Privacy First

100% local: All data on your machine
Works offline: No internet after model download
No telemetry: Your data never leaves your computer

MCP Integration

Connect Roampal to Claude Desktop, Cursor, and other MCP-compatible tools.

Settings → Integrations → Connect → Restart your tool

7 tools available: searchmemory, addtomemorybank, updatememory, archivememory, getcontextinsights, recordresponse, scorememories

Full MCP documentation →

Architecture

┌─────────────────────────────────────────────────────────┐
│                    5-TIER MEMORY                        │
├─────────────┬─────────────┬─────────────┬──────────────┤
│   Books     │   Working   │   History   │   Patterns   │
│ (permanent) │   (24h)     │  (30 days)  │  (permanent) │
├─────────────┴─────────────┴─────────────┴──────────────┤
│                    Memory Bank                          │
│            (permanent user identity/prefs)              │
└─────────────────────────────────────────────────────────┘

Core Technology:

TagCascade Retrieval: Tag-routed search + cross-encoder reranking (ONNX)
Outcome-Based Learning: Memories adapt based on feedback
Sidecar LLM: Background model summarizes exchanges, extracts facts and tags

Architecture deep-dive →

Supported Models

Works with any tool-calling model via Ollama or LM Studio:

| Model | Provider | Parameters | |-------|----------|------------| | Llama 3.x | Meta | 3B - 70B | | Qwen 2.5 | Alibaba | 3B - 72B | | Mistral/Mixtral | Mistral AI | 7B - 8x22B | | GPT-OSS | OpenAI (Apache 2.0) | 20B - 120B |

Documentation

| Document | Description | |----------|-------------| | Architecture | 5-tier memory, retrieval pipeline, technical deep-dive | | Benchmarks | LoCoMo evaluation, TagCascade results | | Release Notes | Latest (v0.3.3): multimodal image input, dynamic capability + context detection, ChromaDB phantom-handling closing issue #8, atomic config writes, Harmony token cleanup |